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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #include <libcfs/libcfs.h>
59 #include <lnet/nidstr.h>
61 #include <lustre/lustre_idl.h>
62 #include <lustre_ha.h>
63 #include <lustre_sec.h>
64 #include <lustre_import.h>
65 #include <lprocfs_status.h>
66 #include <lu_object.h>
67 #include <lustre_req_layout.h>
68 #include <obd_support.h>
69 #include <lustre_ver.h>
71 /* MD flags we _always_ use */
72 #define PTLRPC_MD_OPTIONS 0
75 * Max # of bulk operations in one request.
76 * In order for the client and server to properly negotiate the maximum
77 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
78 * value. The client is free to limit the actual RPC size for any bulk
79 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
80 #define PTLRPC_BULK_OPS_BITS 2
81 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
83 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
84 * should not be used on the server at all. Otherwise, it imposes a
85 * protocol limitation on the maximum RPC size that can be used by any
86 * RPC sent to that server in the future. Instead, the server should
87 * use the negotiated per-client ocd_brw_size to determine the bulk
89 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
92 * Define maxima for bulk I/O.
94 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
95 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
96 * currently supported maximum between peers at connect via ocd_brw_size.
98 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
99 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
100 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
102 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
103 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
104 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
105 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
106 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
107 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
109 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
110 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
111 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
113 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
114 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
116 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
117 # error "PTLRPC_MAX_BRW_SIZE too big"
119 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_PAGES too big"
123 #define PTLRPC_NTHRS_INIT 2
128 * Constants determine how memory is used to buffer incoming service requests.
130 * ?_NBUFS # buffers to allocate when growing the pool
131 * ?_BUFSIZE # bytes in a single request buffer
132 * ?_MAXREQSIZE # maximum request service will receive
134 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
135 * of ?_NBUFS is added to the pool.
137 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
138 * considered full when less than ?_MAXREQSIZE is left in them.
143 * Constants determine how threads are created for ptlrpc service.
145 * ?_NTHRS_INIT # threads to create for each service partition on
146 * initializing. If it's non-affinity service and
147 * there is only one partition, it's the overall #
148 * threads for the service while initializing.
149 * ?_NTHRS_BASE # threads should be created at least for each
150 * ptlrpc partition to keep the service healthy.
151 * It's the low-water mark of threads upper-limit
152 * for each partition.
153 * ?_THR_FACTOR # threads can be added on threads upper-limit for
154 * each CPU core. This factor is only for reference,
155 * we might decrease value of factor if number of cores
156 * per CPT is above a limit.
157 * ?_NTHRS_MAX # overall threads can be created for a service,
158 * it's a soft limit because if service is running
159 * on machine with hundreds of cores and tens of
160 * CPU partitions, we need to guarantee each partition
161 * has ?_NTHRS_BASE threads, which means total threads
162 * will be ?_NTHRS_BASE * number_of_cpts which can
163 * exceed ?_NTHRS_MAX.
167 * #define MDS_NTHRS_INIT 2
168 * #define MDS_NTHRS_BASE 64
169 * #define MDS_NTHRS_FACTOR 8
170 * #define MDS_NTHRS_MAX 1024
173 * ---------------------------------------------------------------------
174 * Server(A) has 16 cores, user configured it to 4 partitions so each
175 * partition has 4 cores, then actual number of service threads on each
177 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
179 * Total number of threads for the service is:
180 * 96 * partitions(4) = 384
183 * ---------------------------------------------------------------------
184 * Server(B) has 32 cores, user configured it to 4 partitions so each
185 * partition has 8 cores, then actual number of service threads on each
187 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
189 * Total number of threads for the service is:
190 * 128 * partitions(4) = 512
193 * ---------------------------------------------------------------------
194 * Server(B) has 96 cores, user configured it to 8 partitions so each
195 * partition has 12 cores, then actual number of service threads on each
197 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
199 * Total number of threads for the service is:
200 * 160 * partitions(8) = 1280
202 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
203 * as upper limit of threads number for each partition:
204 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
207 * ---------------------------------------------------------------------
208 * Server(C) have a thousand of cores and user configured it to 32 partitions
209 * MDS_NTHRS_BASE(64) * 32 = 2048
211 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
212 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
213 * to keep service healthy, so total number of threads will just be 2048.
215 * NB: we don't suggest to choose server with that many cores because backend
216 * filesystem itself, buffer cache, or underlying network stack might
217 * have some SMP scalability issues at that large scale.
219 * If user already has a fat machine with hundreds or thousands of cores,
220 * there are two choices for configuration:
221 * a) create CPU table from subset of all CPUs and run Lustre on
223 * b) bind service threads on a few partitions, see modparameters of
224 * MDS and OSS for details
226 * NB: these calculations (and examples below) are simplified to help
227 * understanding, the real implementation is a little more complex,
228 * please see ptlrpc_server_nthreads_check() for details.
233 * LDLM threads constants:
235 * Given 8 as factor and 24 as base threads number
238 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
241 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
242 * threads for each partition and total threads number will be 112.
245 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
246 * threads for each partition to keep service healthy, so total threads
247 * number should be 24 * 8 = 192.
249 * So with these constants, threads number will be at the similar level
250 * of old versions, unless target machine has over a hundred cores
252 #define LDLM_THR_FACTOR 8
253 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
254 #define LDLM_NTHRS_BASE 24
255 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
257 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
258 #define LDLM_CLIENT_NBUFS 1
259 #define LDLM_SERVER_NBUFS 64
260 #define LDLM_BUFSIZE (8 * 1024)
261 #define LDLM_MAXREQSIZE (5 * 1024)
262 #define LDLM_MAXREPSIZE (1024)
265 * MDS threads constants:
267 * Please see examples in "Thread Constants", MDS threads number will be at
268 * the comparable level of old versions, unless the server has many cores.
270 #ifndef MDS_MAX_THREADS
271 #define MDS_MAX_THREADS 1024
272 #define MDS_MAX_OTHR_THREADS 256
274 #else /* MDS_MAX_THREADS */
275 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
276 #undef MDS_MAX_THREADS
277 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
279 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
282 /* default service */
283 #define MDS_THR_FACTOR 8
284 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
285 #define MDS_NTHRS_MAX MDS_MAX_THREADS
286 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
288 /* read-page service */
289 #define MDS_RDPG_THR_FACTOR 4
290 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
291 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
292 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
294 /* these should be removed when we remove setattr service in the future */
295 #define MDS_SETA_THR_FACTOR 4
296 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
297 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
298 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
300 /* non-affinity threads */
301 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
302 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
307 * Assume file name length = FNAME_MAX = 256 (true for ext3).
308 * path name length = PATH_MAX = 4096
309 * LOV MD size max = EA_MAX = 24 * 2000
310 * (NB: 24 is size of lov_ost_data)
311 * LOV LOGCOOKIE size max = 32 * 2000
312 * (NB: 32 is size of llog_cookie)
313 * symlink: FNAME_MAX + PATH_MAX <- largest
314 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
315 * rename: FNAME_MAX + FNAME_MAX
316 * open: FNAME_MAX + EA_MAX
318 * MDS_MAXREQSIZE ~= 4736 bytes =
319 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
320 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
322 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
323 * except in the open case where there are a large number of OSTs in a LOV.
325 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
326 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
329 * MDS incoming request with LOV EA
330 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
332 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
333 362 + LOV_MAX_STRIPE_COUNT * 24)
335 * MDS outgoing reply with LOV EA
337 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
338 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
340 * but 2.4 or later MDS will never send reply with llog_cookie to any
341 * version client. This macro is defined for server side reply buffer size.
343 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
346 * This is the size of a maximum REINT_SETXATTR request:
348 * lustre_msg 56 (32 + 4 x 5 + 4)
350 * mdt_rec_setxattr 136
352 * name 256 (XATTR_NAME_MAX)
353 * value 65536 (XATTR_SIZE_MAX)
355 #define MDS_EA_MAXREQSIZE 66288
358 * These are the maximum request and reply sizes (rounded up to 1 KB
359 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
361 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
362 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
363 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
366 * The update request includes all of updates from the create, which might
367 * include linkea (4K maxim), together with other updates, we set it to 9K:
368 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
370 #define OUT_MAXREQSIZE (9 * 1024)
371 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
373 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
374 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
378 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
379 * However, we need to allocate a much larger buffer for it because LNet
380 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
381 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
382 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
383 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
384 * utilization is very low.
386 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
387 * reused until all requests fit in it have been processed and released,
388 * which means one long blocked request can prevent the rqbd be reused.
389 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
390 * from LNet with unused 65 KB, buffer utilization will be about 59%.
391 * Please check LU-2432 for details.
393 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
397 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
398 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
399 * extra bytes to each request buffer to improve buffer utilization rate.
401 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
404 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
405 #define FLD_MAXREQSIZE (160)
407 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
408 #define FLD_MAXREPSIZE (152)
409 #define FLD_BUFSIZE (1 << 12)
412 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
414 #define SEQ_MAXREQSIZE (160)
416 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
417 #define SEQ_MAXREPSIZE (152)
418 #define SEQ_BUFSIZE (1 << 12)
420 /** MGS threads must be >= 3, see bug 22458 comment #28 */
421 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
422 #define MGS_NTHRS_MAX 32
425 #define MGS_BUFSIZE (8 * 1024)
426 #define MGS_MAXREQSIZE (7 * 1024)
427 #define MGS_MAXREPSIZE (9 * 1024)
430 * OSS threads constants:
432 * Given 8 as factor and 64 as base threads number
435 * On 8-core server configured to 2 partitions, we will have
436 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
439 * On 32-core machine configured to 4 partitions, we will have
440 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
441 * will be 112 * 4 = 448.
444 * On 64-core machine configured to 4 partitions, we will have
445 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
446 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
447 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
448 * for each partition.
450 * So we can see that with these constants, threads number wil be at the
451 * similar level of old versions, unless the server has many cores.
453 /* depress threads factor for VM with small memory size */
454 #define OSS_THR_FACTOR min_t(int, 8, \
455 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
456 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
457 #define OSS_NTHRS_BASE 64
458 #define OSS_NTHRS_MAX 512
460 /* threads for handling "create" request */
461 #define OSS_CR_THR_FACTOR 1
462 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
463 #define OSS_CR_NTHRS_BASE 8
464 #define OSS_CR_NTHRS_MAX 64
467 * OST_IO_MAXREQSIZE ~=
468 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
469 * DT_MAX_BRW_PAGES * niobuf_remote
471 * - single object with 16 pages is 512 bytes
472 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
473 * - Must be a multiple of 1024
474 * - actual size is about 18K
476 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
477 sizeof(struct ptlrpc_body) + \
478 sizeof(struct obdo) + \
479 sizeof(struct obd_ioobj) + \
480 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
482 * FIEMAP request can be 4K+ for now
484 #define OST_MAXREQSIZE (16 * 1024)
485 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
486 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
488 #define OST_MAXREPSIZE (9 * 1024)
489 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
492 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
493 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
495 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
496 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
498 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
500 /* Macro to hide a typecast. */
501 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
503 struct ptlrpc_replay_async_args {
509 * Structure to single define portal connection.
511 struct ptlrpc_connection {
512 /** linkage for connections hash table */
513 struct hlist_node c_hash;
514 /** Our own lnet nid for this connection */
516 /** Remote side nid for this connection */
517 lnet_process_id_t c_peer;
518 /** UUID of the other side */
519 struct obd_uuid c_remote_uuid;
520 /** reference counter for this connection */
524 /** Client definition for PortalRPC */
525 struct ptlrpc_client {
526 /** What lnet portal does this client send messages to by default */
527 __u32 cli_request_portal;
528 /** What portal do we expect replies on */
529 __u32 cli_reply_portal;
530 /** Name of the client */
534 /** state flags of requests */
535 /* XXX only ones left are those used by the bulk descs as well! */
536 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
537 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
539 #define REQ_MAX_ACK_LOCKS 8
541 union ptlrpc_async_args {
543 * Scratchpad for passing args to completion interpreter. Users
544 * cast to the struct of their choosing, and CLASSERT that this is
545 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
546 * a pointer to it here. The pointer_arg ensures this struct is at
547 * least big enough for that.
549 void *pointer_arg[11];
553 struct ptlrpc_request_set;
554 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
555 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
558 * Definition of request set structure.
559 * Request set is a list of requests (not necessary to the same target) that
560 * once populated with RPCs could be sent in parallel.
561 * There are two kinds of request sets. General purpose and with dedicated
562 * serving thread. Example of the latter is ptlrpcd set.
563 * For general purpose sets once request set started sending it is impossible
564 * to add new requests to such set.
565 * Provides a way to call "completion callbacks" when all requests in the set
568 struct ptlrpc_request_set {
569 atomic_t set_refcount;
570 /** number of in queue requests */
571 atomic_t set_new_count;
572 /** number of uncompleted requests */
573 atomic_t set_remaining;
574 /** wait queue to wait on for request events */
575 wait_queue_head_t set_waitq;
576 wait_queue_head_t *set_wakeup_ptr;
577 /** List of requests in the set */
578 struct list_head set_requests;
580 * List of completion callbacks to be called when the set is completed
581 * This is only used if \a set_interpret is NULL.
582 * Links struct ptlrpc_set_cbdata.
584 struct list_head set_cblist;
585 /** Completion callback, if only one. */
586 set_interpreter_func set_interpret;
587 /** opaq argument passed to completion \a set_interpret callback. */
590 * Lock for \a set_new_requests manipulations
591 * locked so that any old caller can communicate requests to
592 * the set holder who can then fold them into the lock-free set
594 spinlock_t set_new_req_lock;
595 /** List of new yet unsent requests. Only used with ptlrpcd now. */
596 struct list_head set_new_requests;
598 /** rq_status of requests that have been freed already */
600 /** Additional fields used by the flow control extension */
601 /** Maximum number of RPCs in flight */
602 int set_max_inflight;
603 /** Callback function used to generate RPCs */
604 set_producer_func set_producer;
605 /** opaq argument passed to the producer callback */
606 void *set_producer_arg;
610 * Description of a single ptrlrpc_set callback
612 struct ptlrpc_set_cbdata {
613 /** List linkage item */
614 struct list_head psc_item;
615 /** Pointer to interpreting function */
616 set_interpreter_func psc_interpret;
617 /** Opaq argument to pass to the callback */
621 struct ptlrpc_bulk_desc;
622 struct ptlrpc_service_part;
623 struct ptlrpc_service;
626 * ptlrpc callback & work item stuff
628 struct ptlrpc_cb_id {
629 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
630 void *cbid_arg; /* additional arg */
633 /** Maximum number of locks to fit into reply state */
634 #define RS_MAX_LOCKS 8
638 * Structure to define reply state on the server
639 * Reply state holds various reply message information. Also for "difficult"
640 * replies (rep-ack case) we store the state after sending reply and wait
641 * for the client to acknowledge the reception. In these cases locks could be
642 * added to the state for replay/failover consistency guarantees.
644 struct ptlrpc_reply_state {
645 /** Callback description */
646 struct ptlrpc_cb_id rs_cb_id;
647 /** Linkage for list of all reply states in a system */
648 struct list_head rs_list;
649 /** Linkage for list of all reply states on same export */
650 struct list_head rs_exp_list;
651 /** Linkage for list of all reply states for same obd */
652 struct list_head rs_obd_list;
654 struct list_head rs_debug_list;
656 /** A spinlock to protect the reply state flags */
658 /** Reply state flags */
659 unsigned long rs_difficult:1; /* ACK/commit stuff */
660 unsigned long rs_no_ack:1; /* no ACK, even for
661 difficult requests */
662 unsigned long rs_scheduled:1; /* being handled? */
663 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
664 unsigned long rs_handled:1; /* been handled yet? */
665 unsigned long rs_on_net:1; /* reply_out_callback pending? */
666 unsigned long rs_prealloc:1; /* rs from prealloc list */
667 unsigned long rs_committed:1;/* the transaction was committed
668 and the rs was dispatched
669 by ptlrpc_commit_replies */
670 /** Size of the state */
674 /** Transaction number */
678 struct obd_export *rs_export;
679 struct ptlrpc_service_part *rs_svcpt;
680 /** Lnet metadata handle for the reply */
681 lnet_handle_md_t rs_md_h;
682 atomic_t rs_refcount;
684 /** Context for the sevice thread */
685 struct ptlrpc_svc_ctx *rs_svc_ctx;
686 /** Reply buffer (actually sent to the client), encoded if needed */
687 struct lustre_msg *rs_repbuf; /* wrapper */
688 /** Size of the reply buffer */
689 int rs_repbuf_len; /* wrapper buf length */
690 /** Size of the reply message */
691 int rs_repdata_len; /* wrapper msg length */
693 * Actual reply message. Its content is encrupted (if needed) to
694 * produce reply buffer for actual sending. In simple case
695 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
697 struct lustre_msg *rs_msg; /* reply message */
699 /** Number of locks awaiting client ACK */
701 /** Handles of locks awaiting client reply ACK */
702 struct lustre_handle rs_locks[RS_MAX_LOCKS];
703 /** Lock modes of locks in \a rs_locks */
704 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
707 struct ptlrpc_thread;
711 RQ_PHASE_NEW = 0xebc0de00,
712 RQ_PHASE_RPC = 0xebc0de01,
713 RQ_PHASE_BULK = 0xebc0de02,
714 RQ_PHASE_INTERPRET = 0xebc0de03,
715 RQ_PHASE_COMPLETE = 0xebc0de04,
716 RQ_PHASE_UNREGISTERING = 0xebc0de05,
717 RQ_PHASE_UNDEFINED = 0xebc0de06
720 /** Type of request interpreter call-back */
721 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
722 struct ptlrpc_request *req,
724 /** Type of request resend call-back */
725 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
729 * Definition of request pool structure.
730 * The pool is used to store empty preallocated requests for the case
731 * when we would actually need to send something without performing
732 * any allocations (to avoid e.g. OOM).
734 struct ptlrpc_request_pool {
735 /** Locks the list */
737 /** list of ptlrpc_request structs */
738 struct list_head prp_req_list;
739 /** Maximum message size that would fit into a rquest from this pool */
741 /** Function to allocate more requests for this pool */
742 void (*prp_populate)(struct ptlrpc_request_pool *, int);
750 #include <lustre_nrs.h>
753 * Basic request prioritization operations structure.
754 * The whole idea is centered around locks and RPCs that might affect locks.
755 * When a lock is contended we try to give priority to RPCs that might lead
756 * to fastest release of that lock.
757 * Currently only implemented for OSTs only in a way that makes all
758 * IO and truncate RPCs that are coming from a locked region where a lock is
759 * contended a priority over other requests.
761 struct ptlrpc_hpreq_ops {
763 * Check if the lock handle of the given lock is the same as
764 * taken from the request.
766 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
768 * Check if the request is a high priority one.
770 int (*hpreq_check)(struct ptlrpc_request *);
772 * Called after the request has been handled.
774 void (*hpreq_fini)(struct ptlrpc_request *);
777 struct ptlrpc_cli_req {
778 /** For bulk requests on client only: bulk descriptor */
779 struct ptlrpc_bulk_desc *cr_bulk;
780 /** optional time limit for send attempts */
781 cfs_duration_t cr_delay_limit;
782 /** time request was first queued */
783 cfs_time_t cr_queued_time;
784 /** request sent timeval */
785 struct timeval cr_sent_tv;
786 /** time for request really sent out */
788 /** when req reply unlink must finish. */
789 time_t cr_reply_deadline;
790 /** when req bulk unlink must finish. */
791 time_t cr_bulk_deadline;
792 /** Portal to which this request would be sent */
794 /** Portal where to wait for reply and where reply would be sent */
796 /** request resending number */
797 unsigned int cr_resend_nr;
798 /** What was import generation when this request was sent */
800 enum lustre_imp_state cr_send_state;
801 /** Per-request waitq introduced by bug 21938 for recovery waiting */
802 wait_queue_head_t cr_set_waitq;
803 /** Link item for request set lists */
804 struct list_head cr_set_chain;
805 /** link to waited ctx */
806 struct list_head cr_ctx_chain;
808 /** client's half ctx */
809 struct ptlrpc_cli_ctx *cr_cli_ctx;
810 /** Link back to the request set */
811 struct ptlrpc_request_set *cr_set;
812 /** outgoing request MD handle */
813 lnet_handle_md_t cr_req_md_h;
814 /** request-out callback parameter */
815 struct ptlrpc_cb_id cr_req_cbid;
816 /** incoming reply MD handle */
817 lnet_handle_md_t cr_reply_md_h;
818 wait_queue_head_t cr_reply_waitq;
819 /** reply callback parameter */
820 struct ptlrpc_cb_id cr_reply_cbid;
821 /** Async completion handler, called when reply is received */
822 ptlrpc_interpterer_t cr_reply_interp;
823 /** Resend handler, called when request is resend to update RPC data */
824 ptlrpc_resend_cb_t cr_resend_cb;
825 /** Async completion context */
826 union ptlrpc_async_args cr_async_args;
827 /** Opaq data for replay and commit callbacks. */
830 * Commit callback, called when request is committed and about to be
833 void (*cr_commit_cb)(struct ptlrpc_request *);
834 /** Replay callback, called after request is replayed at recovery */
835 void (*cr_replay_cb)(struct ptlrpc_request *);
838 /** client request member alias */
839 /* NB: these alias should NOT be used by any new code, instead they should
840 * be removed step by step to avoid potential abuse */
841 #define rq_bulk rq_cli.cr_bulk
842 #define rq_delay_limit rq_cli.cr_delay_limit
843 #define rq_queued_time rq_cli.cr_queued_time
844 #define rq_sent_tv rq_cli.cr_sent_tv
845 #define rq_real_sent rq_cli.cr_sent_out
846 #define rq_reply_deadline rq_cli.cr_reply_deadline
847 #define rq_bulk_deadline rq_cli.cr_bulk_deadline
848 #define rq_nr_resend rq_cli.cr_resend_nr
849 #define rq_request_portal rq_cli.cr_req_ptl
850 #define rq_reply_portal rq_cli.cr_rep_ptl
851 #define rq_import_generation rq_cli.cr_imp_gen
852 #define rq_send_state rq_cli.cr_send_state
853 #define rq_set_chain rq_cli.cr_set_chain
854 #define rq_ctx_chain rq_cli.cr_ctx_chain
855 #define rq_set rq_cli.cr_set
856 #define rq_set_waitq rq_cli.cr_set_waitq
857 #define rq_cli_ctx rq_cli.cr_cli_ctx
858 #define rq_req_md_h rq_cli.cr_req_md_h
859 #define rq_req_cbid rq_cli.cr_req_cbid
860 #define rq_reply_md_h rq_cli.cr_reply_md_h
861 #define rq_reply_waitq rq_cli.cr_reply_waitq
862 #define rq_reply_cbid rq_cli.cr_reply_cbid
863 #define rq_interpret_reply rq_cli.cr_reply_interp
864 #define rq_resend_cb rq_cli.cr_resend_cb
865 #define rq_async_args rq_cli.cr_async_args
866 #define rq_cb_data rq_cli.cr_cb_data
867 #define rq_commit_cb rq_cli.cr_commit_cb
868 #define rq_replay_cb rq_cli.cr_replay_cb
870 struct ptlrpc_srv_req {
871 /** initial thread servicing this request */
872 struct ptlrpc_thread *sr_svc_thread;
874 * Server side list of incoming unserved requests sorted by arrival
875 * time. Traversed from time to time to notice about to expire
876 * requests and sent back "early replies" to clients to let them
877 * know server is alive and well, just very busy to service their
880 struct list_head sr_timed_list;
881 /** server-side per-export list */
882 struct list_head sr_exp_list;
883 /** server-side history, used for debuging purposes. */
884 struct list_head sr_hist_list;
885 /** history sequence # */
887 /** the index of service's srv_at_array into which request is linked */
891 /** authed uid mapped to */
892 uid_t sr_auth_mapped_uid;
893 /** RPC is generated from what part of Lustre */
894 enum lustre_sec_part sr_sp_from;
895 /** request session context */
896 struct lu_context sr_ses;
900 /** stub for NRS request */
901 struct ptlrpc_nrs_request sr_nrq;
903 /** request arrival time */
904 struct timeval sr_arrival_time;
905 /** server's half ctx */
906 struct ptlrpc_svc_ctx *sr_svc_ctx;
907 /** (server side), pointed directly into req buffer */
908 struct ptlrpc_user_desc *sr_user_desc;
909 /** separated reply state */
910 struct ptlrpc_reply_state *sr_reply_state;
911 /** server-side hp handlers */
912 struct ptlrpc_hpreq_ops *sr_ops;
913 /** incoming request buffer */
914 struct ptlrpc_request_buffer_desc *sr_rqbd;
917 /** server request member alias */
918 /* NB: these alias should NOT be used by any new code, instead they should
919 * be removed step by step to avoid potential abuse */
920 #define rq_svc_thread rq_srv.sr_svc_thread
921 #define rq_timed_list rq_srv.sr_timed_list
922 #define rq_exp_list rq_srv.sr_exp_list
923 #define rq_history_list rq_srv.sr_hist_list
924 #define rq_history_seq rq_srv.sr_hist_seq
925 #define rq_at_index rq_srv.sr_at_index
926 #define rq_auth_uid rq_srv.sr_auth_uid
927 #define rq_auth_mapped_uid rq_srv.sr_auth_mapped_uid
928 #define rq_sp_from rq_srv.sr_sp_from
929 #define rq_session rq_srv.sr_ses
930 #define rq_nrq rq_srv.sr_nrq
931 #define rq_arrival_time rq_srv.sr_arrival_time
932 #define rq_reply_state rq_srv.sr_reply_state
933 #define rq_svc_ctx rq_srv.sr_svc_ctx
934 #define rq_user_desc rq_srv.sr_user_desc
935 #define rq_ops rq_srv.sr_ops
936 #define rq_rqbd rq_srv.sr_rqbd
939 * Represents remote procedure call.
941 * This is a staple structure used by everybody wanting to send a request
944 struct ptlrpc_request {
945 /* Request type: one of PTL_RPC_MSG_* */
947 /** Result of request processing */
950 * Linkage item through which this request is included into
951 * sending/delayed lists on client and into rqbd list on server
953 struct list_head rq_list;
954 /** Lock to protect request flags and some other important bits, like
958 /** client-side flags are serialized by rq_lock @{ */
959 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
960 rq_timedout:1, rq_resend:1, rq_restart:1,
962 * when ->rq_replay is set, request is kept by the client even
963 * after server commits corresponding transaction. This is
964 * used for operations that require sequence of multiple
965 * requests to be replayed. The only example currently is file
966 * open/close. When last request in such a sequence is
967 * committed, ->rq_replay is cleared on all requests in the
971 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
972 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
974 rq_req_unlinked:1, /* unlinked request buffer from lnet */
975 rq_reply_unlinked:1, /* unlinked reply buffer from lnet */
976 rq_memalloc:1, /* req originated from "kswapd" */
978 rq_reply_truncated:1,
979 /** whether the "rq_set" is a valid one */
982 /** do not resend request on -EINPROGRESS */
983 rq_no_retry_einprogress:1,
984 /* allow the req to be sent if the import is in recovery
987 /* bulk request, sent to server, but uncommitted */
991 /** server-side flags @{ */
993 rq_hp:1, /**< high priority RPC */
994 rq_at_linked:1, /**< link into service's srv_at_array */
995 rq_packed_final:1; /**< packed final reply */
998 /** one of RQ_PHASE_* */
999 enum rq_phase rq_phase;
1000 /** one of RQ_PHASE_* to be used next */
1001 enum rq_phase rq_next_phase;
1003 * client-side refcount for SENT race, server-side refcounf
1004 * for multiple replies
1006 atomic_t rq_refcount;
1009 * !rq_truncate : # reply bytes actually received,
1010 * rq_truncate : required repbuf_len for resend
1012 int rq_nob_received;
1013 /** Request length */
1017 /** Pool if request is from preallocated list */
1018 struct ptlrpc_request_pool *rq_pool;
1019 /** Request message - what client sent */
1020 struct lustre_msg *rq_reqmsg;
1021 /** Reply message - server response */
1022 struct lustre_msg *rq_repmsg;
1023 /** Transaction number */
1028 * List item to for replay list. Not yet commited requests get linked
1030 * Also see \a rq_replay comment above.
1031 * It's also link chain on obd_export::exp_req_replay_queue
1033 struct list_head rq_replay_list;
1034 /** non-shared members for client & server request*/
1036 struct ptlrpc_cli_req rq_cli;
1037 struct ptlrpc_srv_req rq_srv;
1040 * security and encryption data
1042 /** description of flavors for client & server */
1043 struct sptlrpc_flavor rq_flvr;
1045 /* client/server security flags */
1047 rq_ctx_init:1, /* context initiation */
1048 rq_ctx_fini:1, /* context destroy */
1049 rq_bulk_read:1, /* request bulk read */
1050 rq_bulk_write:1, /* request bulk write */
1051 /* server authentication flags */
1052 rq_auth_gss:1, /* authenticated by gss */
1053 rq_auth_remote:1, /* authed as remote user */
1054 rq_auth_usr_root:1, /* authed as root */
1055 rq_auth_usr_mdt:1, /* authed as mdt */
1056 rq_auth_usr_ost:1, /* authed as ost */
1057 /* security tfm flags */
1060 /* doesn't expect reply FIXME */
1062 rq_pill_init:1, /* pill initialized */
1063 rq_srv_req:1; /* server request */
1066 /** various buffer pointers */
1067 struct lustre_msg *rq_reqbuf; /**< req wrapper */
1068 char *rq_repbuf; /**< rep buffer */
1069 struct lustre_msg *rq_repdata; /**< rep wrapper msg */
1070 /** only in priv mode */
1071 struct lustre_msg *rq_clrbuf;
1072 int rq_reqbuf_len; /* req wrapper buf len */
1073 int rq_reqdata_len; /* req wrapper msg len */
1074 int rq_repbuf_len; /* rep buffer len */
1075 int rq_repdata_len; /* rep wrapper msg len */
1076 int rq_clrbuf_len; /* only in priv mode */
1077 int rq_clrdata_len; /* only in priv mode */
1079 /** early replies go to offset 0, regular replies go after that */
1080 unsigned int rq_reply_off;
1084 /** Fields that help to see if request and reply were swabbed or not */
1085 __u32 rq_req_swab_mask;
1086 __u32 rq_rep_swab_mask;
1088 /** how many early replies (for stats) */
1090 /** Server-side, export on which request was received */
1091 struct obd_export *rq_export;
1092 /** import where request is being sent */
1093 struct obd_import *rq_import;
1096 /** Peer description (the other side) */
1097 lnet_process_id_t rq_peer;
1099 * service time estimate (secs)
1100 * If the request is not served by this time, it is marked as timed out.
1104 * when request/reply sent (secs), or time when request should be sent
1107 /** when request must finish. */
1109 /** request format description */
1110 struct req_capsule rq_pill;
1114 * Call completion handler for rpc if any, return it's status or original
1115 * rc if there was no handler defined for this request.
1117 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1118 struct ptlrpc_request *req, int rc)
1120 if (req->rq_interpret_reply != NULL) {
1121 req->rq_status = req->rq_interpret_reply(env, req,
1122 &req->rq_async_args,
1124 return req->rq_status;
1132 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1133 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1134 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1135 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1136 struct ptlrpc_nrs_pol_info *info);
1139 * Can the request be moved from the regular NRS head to the high-priority NRS
1140 * head (of the same PTLRPC service partition), if any?
1142 * For a reliable result, this should be checked under svcpt->scp_req lock.
1144 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1146 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1149 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1150 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1151 * to make sure it has not been scheduled yet (analogous to previous
1152 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1154 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1159 * Returns 1 if request buffer at offset \a index was already swabbed
1161 static inline int lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
1163 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1164 return req->rq_req_swab_mask & (1 << index);
1168 * Returns 1 if request reply buffer at offset \a index was already swabbed
1170 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
1172 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1173 return req->rq_rep_swab_mask & (1 << index);
1177 * Returns 1 if request needs to be swabbed into local cpu byteorder
1179 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1181 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1185 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1187 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1189 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1193 * Mark request buffer at offset \a index that it was already swabbed
1195 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
1198 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1199 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1200 req->rq_req_swab_mask |= 1 << index;
1204 * Mark request reply buffer at offset \a index that it was already swabbed
1206 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
1209 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1210 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1211 req->rq_rep_swab_mask |= 1 << index;
1215 * Convert numerical request phase value \a phase into text string description
1217 static inline const char *
1218 ptlrpc_phase2str(enum rq_phase phase)
1227 case RQ_PHASE_INTERPRET:
1229 case RQ_PHASE_COMPLETE:
1231 case RQ_PHASE_UNREGISTERING:
1232 return "Unregistering";
1239 * Convert numerical request phase of the request \a req into text stringi
1242 static inline const char *
1243 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1245 return ptlrpc_phase2str(req->rq_phase);
1249 * Debugging functions and helpers to print request structure into debug log
1252 /* Spare the preprocessor, spoil the bugs. */
1253 #define FLAG(field, str) (field ? str : "")
1255 /** Convert bit flags into a string */
1256 #define DEBUG_REQ_FLAGS(req) \
1257 ptlrpc_rqphase2str(req), \
1258 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1259 FLAG(req->rq_err, "E"), \
1260 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1261 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1262 FLAG(req->rq_no_resend, "N"), \
1263 FLAG(req->rq_waiting, "W"), \
1264 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1265 FLAG(req->rq_committed, "M")
1267 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1269 void _debug_req(struct ptlrpc_request *req,
1270 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1271 __attribute__ ((format (printf, 3, 4)));
1274 * Helper that decides if we need to print request accordig to current debug
1277 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1279 CFS_CHECK_STACK(msgdata, mask, cdls); \
1281 if (((mask) & D_CANTMASK) != 0 || \
1282 ((libcfs_debug & (mask)) != 0 && \
1283 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1284 _debug_req((req), msgdata, fmt, ##a); \
1288 * This is the debug print function you need to use to print request sturucture
1289 * content into lustre debug log.
1290 * for most callers (level is a constant) this is resolved at compile time */
1291 #define DEBUG_REQ(level, req, fmt, args...) \
1293 if ((level) & (D_ERROR | D_WARNING)) { \
1294 static cfs_debug_limit_state_t cdls; \
1295 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1296 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1298 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1299 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1305 * Structure that defines a single page of a bulk transfer
1307 struct ptlrpc_bulk_page {
1308 /** Linkage to list of pages in a bulk */
1309 struct list_head bp_link;
1311 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1314 /** offset within a page */
1316 /** The page itself */
1317 struct page *bp_page;
1320 #define BULK_GET_SOURCE 0
1321 #define BULK_PUT_SINK 1
1322 #define BULK_GET_SINK 2
1323 #define BULK_PUT_SOURCE 3
1326 * Definition of bulk descriptor.
1327 * Bulks are special "Two phase" RPCs where initial request message
1328 * is sent first and it is followed bt a transfer (o receiving) of a large
1329 * amount of data to be settled into pages referenced from the bulk descriptors.
1330 * Bulks transfers (the actual data following the small requests) are done
1331 * on separate LNet portals.
1332 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1333 * Another user is readpage for MDT.
1335 struct ptlrpc_bulk_desc {
1336 /** completed with failure */
1337 unsigned long bd_failure:1;
1338 /** {put,get}{source,sink} */
1339 unsigned long bd_type:2;
1341 unsigned long bd_registered:1;
1342 /** For serialization with callback */
1344 /** Import generation when request for this bulk was sent */
1345 int bd_import_generation;
1346 /** LNet portal for this bulk */
1348 /** Server side - export this bulk created for */
1349 struct obd_export *bd_export;
1350 /** Client side - import this bulk was sent on */
1351 struct obd_import *bd_import;
1352 /** Back pointer to the request */
1353 struct ptlrpc_request *bd_req;
1354 wait_queue_head_t bd_waitq; /* server side only WQ */
1355 int bd_iov_count; /* # entries in bd_iov */
1356 int bd_max_iov; /* allocated size of bd_iov */
1357 int bd_nob; /* # bytes covered */
1358 int bd_nob_transferred; /* # bytes GOT/PUT */
1362 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1363 lnet_nid_t bd_sender; /* stash event::sender */
1364 int bd_md_count; /* # valid entries in bd_mds */
1365 int bd_md_max_brw; /* max entries in bd_mds */
1366 /** array of associated MDs */
1367 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1370 * encrypt iov, size is either 0 or bd_iov_count.
1372 lnet_kiov_t *bd_enc_iov;
1374 lnet_kiov_t bd_iov[0];
1378 SVC_STOPPED = 1 << 0,
1379 SVC_STOPPING = 1 << 1,
1380 SVC_STARTING = 1 << 2,
1381 SVC_RUNNING = 1 << 3,
1383 SVC_SIGNAL = 1 << 5,
1386 #define PTLRPC_THR_NAME_LEN 32
1388 * Definition of server service thread structure
1390 struct ptlrpc_thread {
1392 * List of active threads in svc->srv_threads
1394 struct list_head t_link;
1396 * thread-private data (preallocated memory)
1401 * service thread index, from ptlrpc_start_threads
1405 * service thread pid
1409 * put watchdog in the structure per thread b=14840
1411 struct lc_watchdog *t_watchdog;
1413 * the svc this thread belonged to b=18582
1415 struct ptlrpc_service_part *t_svcpt;
1416 wait_queue_head_t t_ctl_waitq;
1417 struct lu_env *t_env;
1418 char t_name[PTLRPC_THR_NAME_LEN];
1421 static inline int thread_is_init(struct ptlrpc_thread *thread)
1423 return thread->t_flags == 0;
1426 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1428 return !!(thread->t_flags & SVC_STOPPED);
1431 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1433 return !!(thread->t_flags & SVC_STOPPING);
1436 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1438 return !!(thread->t_flags & SVC_STARTING);
1441 static inline int thread_is_running(struct ptlrpc_thread *thread)
1443 return !!(thread->t_flags & SVC_RUNNING);
1446 static inline int thread_is_event(struct ptlrpc_thread *thread)
1448 return !!(thread->t_flags & SVC_EVENT);
1451 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1453 return !!(thread->t_flags & SVC_SIGNAL);
1456 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1458 thread->t_flags &= ~flags;
1461 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1463 thread->t_flags = flags;
1466 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1468 thread->t_flags |= flags;
1471 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1474 if (thread->t_flags & flags) {
1475 thread->t_flags &= ~flags;
1482 * Request buffer descriptor structure.
1483 * This is a structure that contains one posted request buffer for service.
1484 * Once data land into a buffer, event callback creates actual request and
1485 * notifies wakes one of the service threads to process new incoming request.
1486 * More than one request can fit into the buffer.
1488 struct ptlrpc_request_buffer_desc {
1489 /** Link item for rqbds on a service */
1490 struct list_head rqbd_list;
1491 /** History of requests for this buffer */
1492 struct list_head rqbd_reqs;
1493 /** Back pointer to service for which this buffer is registered */
1494 struct ptlrpc_service_part *rqbd_svcpt;
1495 /** LNet descriptor */
1496 lnet_handle_md_t rqbd_md_h;
1498 /** The buffer itself */
1500 struct ptlrpc_cb_id rqbd_cbid;
1502 * This "embedded" request structure is only used for the
1503 * last request to fit into the buffer
1505 struct ptlrpc_request rqbd_req;
1508 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1510 struct ptlrpc_service_ops {
1512 * if non-NULL called during thread creation (ptlrpc_start_thread())
1513 * to initialize service specific per-thread state.
1515 int (*so_thr_init)(struct ptlrpc_thread *thr);
1517 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1518 * destruct state created by ->srv_init().
1520 void (*so_thr_done)(struct ptlrpc_thread *thr);
1522 * Handler function for incoming requests for this service
1524 int (*so_req_handler)(struct ptlrpc_request *req);
1526 * function to determine priority of the request, it's called
1527 * on every new request
1529 int (*so_hpreq_handler)(struct ptlrpc_request *);
1531 * service-specific print fn
1533 void (*so_req_printer)(void *, struct ptlrpc_request *);
1536 #ifndef __cfs_cacheline_aligned
1537 /* NB: put it here for reducing patche dependence */
1538 # define __cfs_cacheline_aligned
1542 * How many high priority requests to serve before serving one normal
1545 #define PTLRPC_SVC_HP_RATIO 10
1548 * Definition of PortalRPC service.
1549 * The service is listening on a particular portal (like tcp port)
1550 * and perform actions for a specific server like IO service for OST
1551 * or general metadata service for MDS.
1553 struct ptlrpc_service {
1554 /** serialize /proc operations */
1555 spinlock_t srv_lock;
1556 /** most often accessed fields */
1557 /** chain thru all services */
1558 struct list_head srv_list;
1559 /** service operations table */
1560 struct ptlrpc_service_ops srv_ops;
1561 /** only statically allocated strings here; we don't clean them */
1563 /** only statically allocated strings here; we don't clean them */
1564 char *srv_thread_name;
1565 /** service thread list */
1566 struct list_head srv_threads;
1567 /** threads # should be created for each partition on initializing */
1568 int srv_nthrs_cpt_init;
1569 /** limit of threads number for each partition */
1570 int srv_nthrs_cpt_limit;
1571 /** Root of /proc dir tree for this service */
1572 struct proc_dir_entry *srv_procroot;
1573 /** Pointer to statistic data for this service */
1574 struct lprocfs_stats *srv_stats;
1575 /** # hp per lp reqs to handle */
1576 int srv_hpreq_ratio;
1577 /** biggest request to receive */
1578 int srv_max_req_size;
1579 /** biggest reply to send */
1580 int srv_max_reply_size;
1581 /** size of individual buffers */
1583 /** # buffers to allocate in 1 group */
1584 int srv_nbuf_per_group;
1585 /** Local portal on which to receive requests */
1586 __u32 srv_req_portal;
1587 /** Portal on the client to send replies to */
1588 __u32 srv_rep_portal;
1590 * Tags for lu_context associated with this thread, see struct
1594 /** soft watchdog timeout multiplier */
1595 int srv_watchdog_factor;
1596 /** under unregister_service */
1597 unsigned srv_is_stopping:1;
1599 /** max # request buffers in history per partition */
1600 int srv_hist_nrqbds_cpt_max;
1601 /** number of CPTs this service bound on */
1603 /** CPTs array this service bound on */
1605 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
1607 /** CPT table this service is running over */
1608 struct cfs_cpt_table *srv_cptable;
1610 * partition data for ptlrpc service
1612 struct ptlrpc_service_part *srv_parts[0];
1616 * Definition of PortalRPC service partition data.
1617 * Although a service only has one instance of it right now, but we
1618 * will have multiple instances very soon (instance per CPT).
1620 * it has four locks:
1622 * serialize operations on rqbd and requests waiting for preprocess
1624 * serialize operations active requests sent to this portal
1626 * serialize adaptive timeout stuff
1628 * serialize operations on RS list (reply states)
1630 * We don't have any use-case to take two or more locks at the same time
1631 * for now, so there is no lock order issue.
1633 struct ptlrpc_service_part {
1634 /** back reference to owner */
1635 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
1636 /* CPT id, reserved */
1638 /** always increasing number */
1640 /** # of starting threads */
1641 int scp_nthrs_starting;
1642 /** # of stopping threads, reserved for shrinking threads */
1643 int scp_nthrs_stopping;
1644 /** # running threads */
1645 int scp_nthrs_running;
1646 /** service threads list */
1647 struct list_head scp_threads;
1650 * serialize the following fields, used for protecting
1651 * rqbd list and incoming requests waiting for preprocess,
1652 * threads starting & stopping are also protected by this lock.
1654 spinlock_t scp_lock __cfs_cacheline_aligned;
1655 /** total # req buffer descs allocated */
1656 int scp_nrqbds_total;
1657 /** # posted request buffers for receiving */
1658 int scp_nrqbds_posted;
1659 /** in progress of allocating rqbd */
1660 int scp_rqbd_allocating;
1661 /** # incoming reqs */
1662 int scp_nreqs_incoming;
1663 /** request buffers to be reposted */
1664 struct list_head scp_rqbd_idle;
1665 /** req buffers receiving */
1666 struct list_head scp_rqbd_posted;
1667 /** incoming reqs */
1668 struct list_head scp_req_incoming;
1669 /** timeout before re-posting reqs, in tick */
1670 cfs_duration_t scp_rqbd_timeout;
1672 * all threads sleep on this. This wait-queue is signalled when new
1673 * incoming request arrives and when difficult reply has to be handled.
1675 wait_queue_head_t scp_waitq;
1677 /** request history */
1678 struct list_head scp_hist_reqs;
1679 /** request buffer history */
1680 struct list_head scp_hist_rqbds;
1681 /** # request buffers in history */
1682 int scp_hist_nrqbds;
1683 /** sequence number for request */
1685 /** highest seq culled from history */
1686 __u64 scp_hist_seq_culled;
1689 * serialize the following fields, used for processing requests
1690 * sent to this portal
1692 spinlock_t scp_req_lock __cfs_cacheline_aligned;
1693 /** # reqs in either of the NRS heads below */
1694 /** # reqs being served */
1695 int scp_nreqs_active;
1696 /** # HPreqs being served */
1697 int scp_nhreqs_active;
1698 /** # hp requests handled */
1701 /** NRS head for regular requests */
1702 struct ptlrpc_nrs scp_nrs_reg;
1703 /** NRS head for HP requests; this is only valid for services that can
1704 * handle HP requests */
1705 struct ptlrpc_nrs *scp_nrs_hp;
1710 * serialize the following fields, used for changes on
1713 spinlock_t scp_at_lock __cfs_cacheline_aligned;
1714 /** estimated rpc service time */
1715 struct adaptive_timeout scp_at_estimate;
1716 /** reqs waiting for replies */
1717 struct ptlrpc_at_array scp_at_array;
1718 /** early reply timer */
1719 struct timer_list scp_at_timer;
1721 cfs_time_t scp_at_checktime;
1722 /** check early replies */
1723 unsigned scp_at_check;
1727 * serialize the following fields, used for processing
1728 * replies for this portal
1730 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
1731 /** all the active replies */
1732 struct list_head scp_rep_active;
1733 /** List of free reply_states */
1734 struct list_head scp_rep_idle;
1735 /** waitq to run, when adding stuff to srv_free_rs_list */
1736 wait_queue_head_t scp_rep_waitq;
1737 /** # 'difficult' replies */
1738 atomic_t scp_nreps_difficult;
1741 #define ptlrpc_service_for_each_part(part, i, svc) \
1743 i < (svc)->srv_ncpts && \
1744 (svc)->srv_parts != NULL && \
1745 ((part) = (svc)->srv_parts[i]) != NULL; i++)
1748 * Declaration of ptlrpcd control structure
1750 struct ptlrpcd_ctl {
1752 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
1754 unsigned long pc_flags;
1756 * Thread lock protecting structure fields.
1762 struct completion pc_starting;
1766 struct completion pc_finishing;
1768 * Thread requests set.
1770 struct ptlrpc_request_set *pc_set;
1772 * Thread name used in kthread_run()
1776 * Environment for request interpreters to run in.
1778 struct lu_env pc_env;
1780 * Index of ptlrpcd thread in the array.
1784 * Number of the ptlrpcd's partners.
1788 * Pointer to the array of partners' ptlrpcd_ctl structure.
1790 struct ptlrpcd_ctl **pc_partners;
1792 * Record the partner index to be processed next.
1797 /* Bits for pc_flags */
1798 enum ptlrpcd_ctl_flags {
1800 * Ptlrpc thread start flag.
1802 LIOD_START = 1 << 0,
1804 * Ptlrpc thread stop flag.
1808 * Ptlrpc thread force flag (only stop force so far).
1809 * This will cause aborting any inflight rpcs handled
1810 * by thread if LIOD_STOP is specified.
1812 LIOD_FORCE = 1 << 2,
1814 * This is a recovery ptlrpc thread.
1816 LIOD_RECOVERY = 1 << 3,
1818 * The ptlrpcd is bound to some CPU core.
1827 * Service compatibility function; the policy is compatible with all services.
1829 * \param[in] svc The service the policy is attempting to register with.
1830 * \param[in] desc The policy descriptor
1832 * \retval true The policy is compatible with the service
1834 * \see ptlrpc_nrs_pol_desc::pd_compat()
1836 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
1837 const struct ptlrpc_nrs_pol_desc *desc)
1843 * Service compatibility function; the policy is compatible with only a specific
1844 * service which is identified by its human-readable name at
1845 * ptlrpc_service::srv_name.
1847 * \param[in] svc The service the policy is attempting to register with.
1848 * \param[in] desc The policy descriptor
1850 * \retval false The policy is not compatible with the service
1851 * \retval true The policy is compatible with the service
1853 * \see ptlrpc_nrs_pol_desc::pd_compat()
1855 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
1856 const struct ptlrpc_nrs_pol_desc *desc)
1858 LASSERT(desc->pd_compat_svc_name != NULL);
1859 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
1864 /* ptlrpc/events.c */
1865 extern lnet_handle_eq_t ptlrpc_eq_h;
1866 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
1867 lnet_process_id_t *peer, lnet_nid_t *self);
1869 * These callbacks are invoked by LNet when something happened to
1873 extern void request_out_callback(lnet_event_t *ev);
1874 extern void reply_in_callback(lnet_event_t *ev);
1875 extern void client_bulk_callback(lnet_event_t *ev);
1876 extern void request_in_callback(lnet_event_t *ev);
1877 extern void reply_out_callback(lnet_event_t *ev);
1878 #ifdef HAVE_SERVER_SUPPORT
1879 extern void server_bulk_callback(lnet_event_t *ev);
1883 /* ptlrpc/connection.c */
1884 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
1886 struct obd_uuid *uuid);
1887 int ptlrpc_connection_put(struct ptlrpc_connection *c);
1888 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
1889 int ptlrpc_connection_init(void);
1890 void ptlrpc_connection_fini(void);
1891 extern lnet_pid_t ptl_get_pid(void);
1893 /* ptlrpc/niobuf.c */
1895 * Actual interfacing with LNet to put/get/register/unregister stuff
1898 #ifdef HAVE_SERVER_SUPPORT
1899 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
1900 unsigned npages, unsigned max_brw,
1901 unsigned type, unsigned portal);
1902 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
1903 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
1905 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
1909 LASSERT(desc != NULL);
1911 spin_lock(&desc->bd_lock);
1912 rc = desc->bd_md_count;
1913 spin_unlock(&desc->bd_lock);
1918 int ptlrpc_register_bulk(struct ptlrpc_request *req);
1919 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
1921 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
1923 struct ptlrpc_bulk_desc *desc;
1926 LASSERT(req != NULL);
1927 desc = req->rq_bulk;
1929 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
1930 req->rq_bulk_deadline > cfs_time_current_sec())
1936 spin_lock(&desc->bd_lock);
1937 rc = desc->bd_md_count;
1938 spin_unlock(&desc->bd_lock);
1942 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
1943 #define PTLRPC_REPLY_EARLY 0x02
1944 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
1945 int ptlrpc_reply(struct ptlrpc_request *req);
1946 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
1947 int ptlrpc_error(struct ptlrpc_request *req);
1948 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
1949 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
1950 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
1953 /* ptlrpc/client.c */
1955 * Client-side portals API. Everything to send requests, receive replies,
1956 * request queues, request management, etc.
1959 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
1961 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
1962 struct ptlrpc_client *);
1963 void ptlrpc_cleanup_client(struct obd_import *imp);
1964 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
1966 int ptlrpc_queue_wait(struct ptlrpc_request *req);
1967 int ptlrpc_replay_req(struct ptlrpc_request *req);
1968 void ptlrpc_restart_req(struct ptlrpc_request *req);
1969 void ptlrpc_abort_inflight(struct obd_import *imp);
1970 void ptlrpc_cleanup_imp(struct obd_import *imp);
1971 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
1973 struct ptlrpc_request_set *ptlrpc_prep_set(void);
1974 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1976 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1977 set_interpreter_func fn, void *data);
1978 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
1979 int ptlrpc_set_wait(struct ptlrpc_request_set *);
1980 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
1981 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
1982 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
1984 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
1985 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
1987 struct ptlrpc_request_pool *
1988 ptlrpc_init_rq_pool(int, int,
1989 void (*populate_pool)(struct ptlrpc_request_pool *, int));
1991 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
1992 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1993 const struct req_format *format);
1994 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
1995 struct ptlrpc_request_pool *,
1996 const struct req_format *format);
1997 void ptlrpc_request_free(struct ptlrpc_request *request);
1998 int ptlrpc_request_pack(struct ptlrpc_request *request,
1999 __u32 version, int opcode);
2000 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2001 const struct req_format *format,
2002 __u32 version, int opcode);
2003 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2004 __u32 version, int opcode, char **bufs,
2005 struct ptlrpc_cli_ctx *ctx);
2006 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2007 int opcode, int count, __u32 *lengths,
2009 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2010 __u32 version, int opcode,
2011 int count, __u32 *lengths, char **bufs,
2012 struct ptlrpc_request_pool *pool);
2013 void ptlrpc_req_finished(struct ptlrpc_request *request);
2014 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2015 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2016 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2017 unsigned npages, unsigned max_brw,
2018 unsigned type, unsigned portal);
2019 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2020 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2022 __ptlrpc_free_bulk(bulk, 1);
2024 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2026 __ptlrpc_free_bulk(bulk, 0);
2028 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2029 struct page *page, int pageoffset, int len, int);
2030 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2031 struct page *page, int pageoffset,
2034 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2037 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2038 struct page *page, int pageoffset,
2041 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2044 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2045 struct obd_import *imp);
2046 __u64 ptlrpc_next_xid(void);
2047 __u64 ptlrpc_sample_next_xid(void);
2048 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2050 /* Set of routines to run a function in ptlrpcd context */
2051 void *ptlrpcd_alloc_work(struct obd_import *imp,
2052 int (*cb)(const struct lu_env *, void *), void *data);
2053 void ptlrpcd_destroy_work(void *handler);
2054 int ptlrpcd_queue_work(void *handler);
2057 struct ptlrpc_service_buf_conf {
2058 /* nbufs is buffers # to allocate when growing the pool */
2059 unsigned int bc_nbufs;
2060 /* buffer size to post */
2061 unsigned int bc_buf_size;
2062 /* portal to listed for requests on */
2063 unsigned int bc_req_portal;
2064 /* portal of where to send replies to */
2065 unsigned int bc_rep_portal;
2066 /* maximum request size to be accepted for this service */
2067 unsigned int bc_req_max_size;
2068 /* maximum reply size this service can ever send */
2069 unsigned int bc_rep_max_size;
2072 struct ptlrpc_service_thr_conf {
2073 /* threadname should be 8 characters or less - 6 will be added on */
2075 /* threads increasing factor for each CPU */
2076 unsigned int tc_thr_factor;
2077 /* service threads # to start on each partition while initializing */
2078 unsigned int tc_nthrs_init;
2080 * low water of threads # upper-limit on each partition while running,
2081 * service availability may be impacted if threads number is lower
2082 * than this value. It can be ZERO if the service doesn't require
2083 * CPU affinity or there is only one partition.
2085 unsigned int tc_nthrs_base;
2086 /* "soft" limit for total threads number */
2087 unsigned int tc_nthrs_max;
2088 /* user specified threads number, it will be validated due to
2089 * other members of this structure. */
2090 unsigned int tc_nthrs_user;
2091 /* set NUMA node affinity for service threads */
2092 unsigned int tc_cpu_affinity;
2093 /* Tags for lu_context associated with service thread */
2097 struct ptlrpc_service_cpt_conf {
2098 struct cfs_cpt_table *cc_cptable;
2099 /* string pattern to describe CPTs for a service */
2103 struct ptlrpc_service_conf {
2106 /* soft watchdog timeout multiplifier to print stuck service traces */
2107 unsigned int psc_watchdog_factor;
2108 /* buffer information */
2109 struct ptlrpc_service_buf_conf psc_buf;
2110 /* thread information */
2111 struct ptlrpc_service_thr_conf psc_thr;
2112 /* CPU partition information */
2113 struct ptlrpc_service_cpt_conf psc_cpt;
2114 /* function table */
2115 struct ptlrpc_service_ops psc_ops;
2118 /* ptlrpc/service.c */
2120 * Server-side services API. Register/unregister service, request state
2121 * management, service thread management
2125 void ptlrpc_save_lock(struct ptlrpc_request *req,
2126 struct lustre_handle *lock, int mode, int no_ack);
2127 void ptlrpc_commit_replies(struct obd_export *exp);
2128 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2129 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2130 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2131 struct ptlrpc_service *ptlrpc_register_service(
2132 struct ptlrpc_service_conf *conf,
2133 struct proc_dir_entry *proc_entry);
2134 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2136 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2137 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2138 int liblustre_check_services(void *arg);
2139 void ptlrpc_daemonize(char *name);
2140 int ptlrpc_service_health_check(struct ptlrpc_service *);
2141 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2142 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2143 struct obd_export *export);
2144 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay);
2146 int ptlrpc_hr_init(void);
2147 void ptlrpc_hr_fini(void);
2151 /* ptlrpc/import.c */
2156 int ptlrpc_connect_import(struct obd_import *imp);
2157 int ptlrpc_init_import(struct obd_import *imp);
2158 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2159 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2160 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2163 /* ptlrpc/pack_generic.c */
2164 int ptlrpc_reconnect_import(struct obd_import *imp);
2168 * ptlrpc msg buffer and swab interface
2172 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2174 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2176 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2177 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2179 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2180 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2182 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2183 __u32 *lens, char **bufs);
2184 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2186 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2187 __u32 *lens, char **bufs, int flags);
2188 #define LPRFL_EARLY_REPLY 1
2189 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2190 char **bufs, int flags);
2191 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2192 unsigned int newlen, int move_data);
2193 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2194 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2195 __u32 lustre_msg_hdr_size(__u32 magic, __u32 count);
2196 __u32 lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2197 __u32 lustre_msg_size_v2(int count, __u32 *lengths);
2198 __u32 lustre_packed_msg_size(struct lustre_msg *msg);
2199 __u32 lustre_msg_early_size(void);
2200 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, __u32 n, __u32 min_size);
2201 void *lustre_msg_buf(struct lustre_msg *m, __u32 n, __u32 minlen);
2202 __u32 lustre_msg_buflen(struct lustre_msg *m, __u32 n);
2203 void lustre_msg_set_buflen(struct lustre_msg *m, __u32 n, __u32 len);
2204 __u32 lustre_msg_bufcount(struct lustre_msg *m);
2205 char *lustre_msg_string(struct lustre_msg *m, __u32 n, __u32 max_len);
2206 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2207 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2208 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2209 void lustre_msg_add_flags(struct lustre_msg *msg, __u32 flags);
2210 void lustre_msg_set_flags(struct lustre_msg *msg, __u32 flags);
2211 void lustre_msg_clear_flags(struct lustre_msg *msg, __u32 flags);
2212 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2213 void lustre_msg_add_op_flags(struct lustre_msg *msg, __u32 flags);
2214 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2215 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2216 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2217 void lustre_msg_add_version(struct lustre_msg *msg, __u32 version);
2218 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2219 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2220 __u16 lustre_msg_get_tag(struct lustre_msg *msg);
2221 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2222 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2223 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2224 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2225 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2226 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2227 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2228 int lustre_msg_get_status(struct lustre_msg *msg);
2229 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2230 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2231 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2232 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2233 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2234 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2235 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 53, 0)
2236 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2238 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2240 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2241 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2242 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2243 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2244 void lustre_msg_set_tag(struct lustre_msg *msg, __u16 tag);
2245 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2246 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2247 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2248 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2249 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2250 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2251 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2252 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2253 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2254 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2255 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2258 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2259 unsigned int newlen, int move_data)
2261 LASSERT(req->rq_reply_state);
2262 LASSERT(req->rq_repmsg);
2263 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2267 #ifdef LUSTRE_TRANSLATE_ERRNOS
2269 static inline int ptlrpc_status_hton(int h)
2272 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2273 * ELDLM_LOCK_ABORTED, etc.
2276 return -lustre_errno_hton(-h);
2281 static inline int ptlrpc_status_ntoh(int n)
2284 * See the comment in ptlrpc_status_hton().
2287 return -lustre_errno_ntoh(-n);
2294 #define ptlrpc_status_hton(h) (h)
2295 #define ptlrpc_status_ntoh(n) (n)
2300 /** Change request phase of \a req to \a new_phase */
2302 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2304 if (req->rq_phase == new_phase)
2307 if (new_phase == RQ_PHASE_UNREGISTERING) {
2308 req->rq_next_phase = req->rq_phase;
2310 atomic_inc(&req->rq_import->imp_unregistering);
2313 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2315 atomic_dec(&req->rq_import->imp_unregistering);
2318 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2319 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2321 req->rq_phase = new_phase;
2325 * Returns true if request \a req got early reply and hard deadline is not met
2328 ptlrpc_client_early(struct ptlrpc_request *req)
2330 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2331 req->rq_reply_deadline > cfs_time_current_sec())
2333 return req->rq_early;
2337 * Returns true if we got real reply from server for this request
2340 ptlrpc_client_replied(struct ptlrpc_request *req)
2342 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2343 req->rq_reply_deadline > cfs_time_current_sec())
2345 return req->rq_replied;
2348 /** Returns true if request \a req is in process of receiving server reply */
2350 ptlrpc_client_recv(struct ptlrpc_request *req)
2352 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2353 req->rq_reply_deadline > cfs_time_current_sec())
2355 return req->rq_receiving_reply;
2359 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2363 spin_lock(&req->rq_lock);
2364 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2365 req->rq_reply_deadline > cfs_time_current_sec()) {
2366 spin_unlock(&req->rq_lock);
2369 rc = !req->rq_req_unlinked || !req->rq_reply_unlinked ||
2370 req->rq_receiving_reply;
2371 spin_unlock(&req->rq_lock);
2376 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2378 if (req->rq_set == NULL)
2379 wake_up(&req->rq_reply_waitq);
2381 wake_up(&req->rq_set->set_waitq);
2385 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2387 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2388 atomic_inc(&rs->rs_refcount);
2392 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2394 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2395 if (atomic_dec_and_test(&rs->rs_refcount))
2396 lustre_free_reply_state(rs);
2399 /* Should only be called once per req */
2400 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2402 if (req->rq_reply_state == NULL)
2403 return; /* shouldn't occur */
2404 ptlrpc_rs_decref(req->rq_reply_state);
2405 req->rq_reply_state = NULL;
2406 req->rq_repmsg = NULL;
2409 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2411 return lustre_msg_get_magic(req->rq_reqmsg);
2414 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2416 switch (req->rq_reqmsg->lm_magic) {
2417 case LUSTRE_MSG_MAGIC_V2:
2418 return req->rq_reqmsg->lm_repsize;
2420 LASSERTF(0, "incorrect message magic: %08x\n",
2421 req->rq_reqmsg->lm_magic);
2426 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2428 if (req->rq_delay_limit != 0 &&
2429 cfs_time_before(cfs_time_add(req->rq_queued_time,
2430 cfs_time_seconds(req->rq_delay_limit)),
2431 cfs_time_current())) {
2437 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2439 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2440 spin_lock(&req->rq_lock);
2441 req->rq_no_resend = 1;
2442 spin_unlock(&req->rq_lock);
2444 return req->rq_no_resend;
2448 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2450 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2452 return svcpt->scp_service->srv_watchdog_factor *
2453 max_t(int, at, obd_timeout);
2456 static inline struct ptlrpc_service *
2457 ptlrpc_req2svc(struct ptlrpc_request *req)
2459 LASSERT(req->rq_rqbd != NULL);
2460 return req->rq_rqbd->rqbd_svcpt->scp_service;
2463 /* ldlm/ldlm_lib.c */
2465 * Target client logic
2468 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2469 int client_obd_cleanup(struct obd_device *obddev);
2470 int client_connect_import(const struct lu_env *env,
2471 struct obd_export **exp, struct obd_device *obd,
2472 struct obd_uuid *cluuid, struct obd_connect_data *,
2474 int client_disconnect_export(struct obd_export *exp);
2475 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2477 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2478 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2479 struct obd_uuid *uuid);
2480 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2481 void client_destroy_import(struct obd_import *imp);
2484 #ifdef HAVE_SERVER_SUPPORT
2485 int server_disconnect_export(struct obd_export *exp);
2488 /* ptlrpc/pinger.c */
2490 * Pinger API (client side only)
2493 enum timeout_event {
2496 struct timeout_item;
2497 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2498 int ptlrpc_pinger_add_import(struct obd_import *imp);
2499 int ptlrpc_pinger_del_import(struct obd_import *imp);
2500 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
2501 timeout_cb_t cb, void *data,
2502 struct list_head *obd_list);
2503 int ptlrpc_del_timeout_client(struct list_head *obd_list,
2504 enum timeout_event event);
2505 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
2506 int ptlrpc_obd_ping(struct obd_device *obd);
2507 void ping_evictor_start(void);
2508 void ping_evictor_stop(void);
2509 void ptlrpc_pinger_ir_up(void);
2510 void ptlrpc_pinger_ir_down(void);
2512 int ptlrpc_pinger_suppress_pings(void);
2514 /* ptlrpc daemon bind policy */
2516 /* all ptlrpcd threads are free mode */
2517 PDB_POLICY_NONE = 1,
2518 /* all ptlrpcd threads are bound mode */
2519 PDB_POLICY_FULL = 2,
2520 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
2521 PDB_POLICY_PAIR = 3,
2522 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
2523 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
2524 * If kernel supports NUMA, pthrpcd threads are binded and
2525 * grouped by NUMA node */
2526 PDB_POLICY_NEIGHBOR = 4,
2529 /* ptlrpc daemon load policy
2530 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
2531 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
2532 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
2533 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
2534 * depends on which is scheduled firstly, to accelerate the RPC processing. */
2536 /* on the same CPU core as the caller */
2537 PDL_POLICY_SAME = 1,
2538 /* within the same CPU partition, but not the same core as the caller */
2539 PDL_POLICY_LOCAL = 2,
2540 /* round-robin on all CPU cores, but not the same core as the caller */
2541 PDL_POLICY_ROUND = 3,
2542 /* the specified CPU core is preferred, but not enforced */
2543 PDL_POLICY_PREFERRED = 4,
2546 /* ptlrpc/ptlrpcd.c */
2547 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2548 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2549 void ptlrpcd_wake(struct ptlrpc_request *req);
2550 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
2551 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2552 int ptlrpcd_addref(void);
2553 void ptlrpcd_decref(void);
2555 /* ptlrpc/lproc_ptlrpc.c */
2557 * procfs output related functions
2560 const char* ll_opcode2str(__u32 opcode);
2561 #ifdef CONFIG_PROC_FS
2562 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2563 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2564 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2566 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2567 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2568 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2572 /* ptlrpc/llog_server.c */
2573 int llog_origin_handle_open(struct ptlrpc_request *req);
2574 int llog_origin_handle_destroy(struct ptlrpc_request *req);
2575 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
2576 int llog_origin_handle_next_block(struct ptlrpc_request *req);
2577 int llog_origin_handle_read_header(struct ptlrpc_request *req);
2578 int llog_origin_handle_close(struct ptlrpc_request *req);
2580 /* ptlrpc/llog_client.c */
2581 extern struct llog_operations llog_client_ops;