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, 2013, 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/lnet.h>
60 #include <lustre/lustre_idl.h>
61 #include <lustre_ha.h>
62 #include <lustre_sec.h>
63 #include <lustre_import.h>
64 #include <lprocfs_status.h>
65 #include <lu_object.h>
66 #include <lustre_req_layout.h>
67 #include <obd_support.h>
68 #include <lustre_ver.h>
70 /* MD flags we _always_ use */
71 #define PTLRPC_MD_OPTIONS 0
74 * Max # of bulk operations in one request.
75 * In order for the client and server to properly negotiate the maximum
76 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
77 * value. The client is free to limit the actual RPC size for any bulk
78 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
79 #define PTLRPC_BULK_OPS_BITS 2
80 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
82 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
83 * should not be used on the server at all. Otherwise, it imposes a
84 * protocol limitation on the maximum RPC size that can be used by any
85 * RPC sent to that server in the future. Instead, the server should
86 * use the negotiated per-client ocd_brw_size to determine the bulk
88 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
91 * Define maxima for bulk I/O.
93 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
94 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
95 * currently supported maximum between peers at connect via ocd_brw_size.
97 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
98 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
99 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
101 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
102 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
103 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
104 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
105 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
106 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
108 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
109 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
110 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
112 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
113 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
115 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
116 # error "PTLRPC_MAX_BRW_SIZE too big"
118 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
119 # error "PTLRPC_MAX_BRW_PAGES too big"
122 #define PTLRPC_NTHRS_INIT 2
127 * Constants determine how memory is used to buffer incoming service requests.
129 * ?_NBUFS # buffers to allocate when growing the pool
130 * ?_BUFSIZE # bytes in a single request buffer
131 * ?_MAXREQSIZE # maximum request service will receive
133 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
134 * of ?_NBUFS is added to the pool.
136 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
137 * considered full when less than ?_MAXREQSIZE is left in them.
142 * Constants determine how threads are created for ptlrpc service.
144 * ?_NTHRS_INIT # threads to create for each service partition on
145 * initializing. If it's non-affinity service and
146 * there is only one partition, it's the overall #
147 * threads for the service while initializing.
148 * ?_NTHRS_BASE # threads should be created at least for each
149 * ptlrpc partition to keep the service healthy.
150 * It's the low-water mark of threads upper-limit
151 * for each partition.
152 * ?_THR_FACTOR # threads can be added on threads upper-limit for
153 * each CPU core. This factor is only for reference,
154 * we might decrease value of factor if number of cores
155 * per CPT is above a limit.
156 * ?_NTHRS_MAX # overall threads can be created for a service,
157 * it's a soft limit because if service is running
158 * on machine with hundreds of cores and tens of
159 * CPU partitions, we need to guarantee each partition
160 * has ?_NTHRS_BASE threads, which means total threads
161 * will be ?_NTHRS_BASE * number_of_cpts which can
162 * exceed ?_NTHRS_MAX.
166 * #define MDS_NTHRS_INIT 2
167 * #define MDS_NTHRS_BASE 64
168 * #define MDS_NTHRS_FACTOR 8
169 * #define MDS_NTHRS_MAX 1024
172 * ---------------------------------------------------------------------
173 * Server(A) has 16 cores, user configured it to 4 partitions so each
174 * partition has 4 cores, then actual number of service threads on each
176 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
178 * Total number of threads for the service is:
179 * 96 * partitions(4) = 384
182 * ---------------------------------------------------------------------
183 * Server(B) has 32 cores, user configured it to 4 partitions so each
184 * partition has 8 cores, then actual number of service threads on each
186 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
188 * Total number of threads for the service is:
189 * 128 * partitions(4) = 512
192 * ---------------------------------------------------------------------
193 * Server(B) has 96 cores, user configured it to 8 partitions so each
194 * partition has 12 cores, then actual number of service threads on each
196 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
198 * Total number of threads for the service is:
199 * 160 * partitions(8) = 1280
201 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
202 * as upper limit of threads number for each partition:
203 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
206 * ---------------------------------------------------------------------
207 * Server(C) have a thousand of cores and user configured it to 32 partitions
208 * MDS_NTHRS_BASE(64) * 32 = 2048
210 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
211 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
212 * to keep service healthy, so total number of threads will just be 2048.
214 * NB: we don't suggest to choose server with that many cores because backend
215 * filesystem itself, buffer cache, or underlying network stack might
216 * have some SMP scalability issues at that large scale.
218 * If user already has a fat machine with hundreds or thousands of cores,
219 * there are two choices for configuration:
220 * a) create CPU table from subset of all CPUs and run Lustre on
222 * b) bind service threads on a few partitions, see modparameters of
223 * MDS and OSS for details
225 * NB: these calculations (and examples below) are simplified to help
226 * understanding, the real implementation is a little more complex,
227 * please see ptlrpc_server_nthreads_check() for details.
232 * LDLM threads constants:
234 * Given 8 as factor and 24 as base threads number
237 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
240 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
241 * threads for each partition and total threads number will be 112.
244 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
245 * threads for each partition to keep service healthy, so total threads
246 * number should be 24 * 8 = 192.
248 * So with these constants, threads number will be at the similar level
249 * of old versions, unless target machine has over a hundred cores
251 #define LDLM_THR_FACTOR 8
252 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
253 #define LDLM_NTHRS_BASE 24
254 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
256 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
257 #define LDLM_CLIENT_NBUFS 1
258 #define LDLM_SERVER_NBUFS 64
259 #define LDLM_BUFSIZE (8 * 1024)
260 #define LDLM_MAXREQSIZE (5 * 1024)
261 #define LDLM_MAXREPSIZE (1024)
264 * MDS threads constants:
266 * Please see examples in "Thread Constants", MDS threads number will be at
267 * the comparable level of old versions, unless the server has many cores.
269 #ifndef MDS_MAX_THREADS
270 #define MDS_MAX_THREADS 1024
271 #define MDS_MAX_OTHR_THREADS 256
273 #else /* MDS_MAX_THREADS */
274 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
275 #undef MDS_MAX_THREADS
276 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
278 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
281 /* default service */
282 #define MDS_THR_FACTOR 8
283 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
284 #define MDS_NTHRS_MAX MDS_MAX_THREADS
285 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
287 /* read-page service */
288 #define MDS_RDPG_THR_FACTOR 4
289 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
290 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
291 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
293 /* these should be removed when we remove setattr service in the future */
294 #define MDS_SETA_THR_FACTOR 4
295 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
296 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
297 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
299 /* non-affinity threads */
300 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
301 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
306 * Assume file name length = FNAME_MAX = 256 (true for ext3).
307 * path name length = PATH_MAX = 4096
308 * LOV MD size max = EA_MAX = 24 * 2000
309 * (NB: 24 is size of lov_ost_data)
310 * LOV LOGCOOKIE size max = 32 * 2000
311 * (NB: 32 is size of llog_cookie)
312 * symlink: FNAME_MAX + PATH_MAX <- largest
313 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
314 * rename: FNAME_MAX + FNAME_MAX
315 * open: FNAME_MAX + EA_MAX
317 * MDS_MAXREQSIZE ~= 4736 bytes =
318 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
319 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
321 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
322 * except in the open case where there are a large number of OSTs in a LOV.
324 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
325 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
328 * MDS incoming request with LOV EA
329 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
331 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
332 362 + LOV_MAX_STRIPE_COUNT * 24)
334 * MDS outgoing reply with LOV EA
336 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
337 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
339 * but 2.4 or later MDS will never send reply with llog_cookie to any
340 * version client. This macro is defined for server side reply buffer size.
342 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
345 * This is the size of a maximum REINT_SETXATTR request:
347 * lustre_msg 56 (32 + 4 x 5 + 4)
349 * mdt_rec_setxattr 136
351 * name 256 (XATTR_NAME_MAX)
352 * value 65536 (XATTR_SIZE_MAX)
354 #define MDS_EA_MAXREQSIZE 66288
357 * These are the maximum request and reply sizes (rounded up to 1 KB
358 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
360 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
361 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
362 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
365 * The update request includes all of updates from the create, which might
366 * include linkea (4K maxim), together with other updates, we set it to 9K:
367 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
369 #define OUT_MAXREQSIZE (9 * 1024)
370 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
372 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
373 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
377 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
378 * However, we need to allocate a much larger buffer for it because LNet
379 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
380 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
381 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
382 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
383 * utilization is very low.
385 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
386 * reused until all requests fit in it have been processed and released,
387 * which means one long blocked request can prevent the rqbd be reused.
388 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
389 * from LNet with unused 65 KB, buffer utilization will be about 59%.
390 * Please check LU-2432 for details.
392 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
396 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
397 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
398 * extra bytes to each request buffer to improve buffer utilization rate.
400 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
403 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
404 #define FLD_MAXREQSIZE (160)
406 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
407 #define FLD_MAXREPSIZE (152)
408 #define FLD_BUFSIZE (1 << 12)
411 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
413 #define SEQ_MAXREQSIZE (160)
415 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
416 #define SEQ_MAXREPSIZE (152)
417 #define SEQ_BUFSIZE (1 << 12)
419 /** MGS threads must be >= 3, see bug 22458 comment #28 */
420 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
421 #define MGS_NTHRS_MAX 32
424 #define MGS_BUFSIZE (8 * 1024)
425 #define MGS_MAXREQSIZE (7 * 1024)
426 #define MGS_MAXREPSIZE (9 * 1024)
429 * OSS threads constants:
431 * Given 8 as factor and 64 as base threads number
434 * On 8-core server configured to 2 partitions, we will have
435 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
438 * On 32-core machine configured to 4 partitions, we will have
439 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
440 * will be 112 * 4 = 448.
443 * On 64-core machine configured to 4 partitions, we will have
444 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
445 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
446 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
447 * for each partition.
449 * So we can see that with these constants, threads number wil be at the
450 * similar level of old versions, unless the server has many cores.
452 /* depress threads factor for VM with small memory size */
453 #define OSS_THR_FACTOR min_t(int, 8, \
454 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
455 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
456 #define OSS_NTHRS_BASE 64
457 #define OSS_NTHRS_MAX 512
459 /* threads for handling "create" request */
460 #define OSS_CR_THR_FACTOR 1
461 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
462 #define OSS_CR_NTHRS_BASE 8
463 #define OSS_CR_NTHRS_MAX 64
466 * OST_IO_MAXREQSIZE ~=
467 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
468 * DT_MAX_BRW_PAGES * niobuf_remote
470 * - single object with 16 pages is 512 bytes
471 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
472 * - Must be a multiple of 1024
473 * - actual size is about 18K
475 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
476 sizeof(struct ptlrpc_body) + \
477 sizeof(struct obdo) + \
478 sizeof(struct obd_ioobj) + \
479 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
481 * FIEMAP request can be 4K+ for now
483 #define OST_MAXREQSIZE (16 * 1024)
484 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
485 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
487 #define OST_MAXREPSIZE (9 * 1024)
488 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
491 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
492 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
494 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
495 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
497 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
499 /* Macro to hide a typecast. */
500 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
503 * Structure to single define portal connection.
505 struct ptlrpc_connection {
506 /** linkage for connections hash table */
507 struct hlist_node c_hash;
508 /** Our own lnet nid for this connection */
510 /** Remote side nid for this connection */
511 lnet_process_id_t c_peer;
512 /** UUID of the other side */
513 struct obd_uuid c_remote_uuid;
514 /** reference counter for this connection */
518 /** Client definition for PortalRPC */
519 struct ptlrpc_client {
520 /** What lnet portal does this client send messages to by default */
521 __u32 cli_request_portal;
522 /** What portal do we expect replies on */
523 __u32 cli_reply_portal;
524 /** Name of the client */
528 /** state flags of requests */
529 /* XXX only ones left are those used by the bulk descs as well! */
530 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
531 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
533 #define REQ_MAX_ACK_LOCKS 8
535 union ptlrpc_async_args {
537 * Scratchpad for passing args to completion interpreter. Users
538 * cast to the struct of their choosing, and CLASSERT that this is
539 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
540 * a pointer to it here. The pointer_arg ensures this struct is at
541 * least big enough for that.
543 void *pointer_arg[11];
547 struct ptlrpc_request_set;
548 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
549 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
552 * Definition of request set structure.
553 * Request set is a list of requests (not necessary to the same target) that
554 * once populated with RPCs could be sent in parallel.
555 * There are two kinds of request sets. General purpose and with dedicated
556 * serving thread. Example of the latter is ptlrpcd set.
557 * For general purpose sets once request set started sending it is impossible
558 * to add new requests to such set.
559 * Provides a way to call "completion callbacks" when all requests in the set
562 struct ptlrpc_request_set {
563 atomic_t set_refcount;
564 /** number of in queue requests */
565 atomic_t set_new_count;
566 /** number of uncompleted requests */
567 atomic_t set_remaining;
568 /** wait queue to wait on for request events */
569 wait_queue_head_t set_waitq;
570 wait_queue_head_t *set_wakeup_ptr;
571 /** List of requests in the set */
572 struct list_head set_requests;
574 * List of completion callbacks to be called when the set is completed
575 * This is only used if \a set_interpret is NULL.
576 * Links struct ptlrpc_set_cbdata.
578 struct list_head set_cblist;
579 /** Completion callback, if only one. */
580 set_interpreter_func set_interpret;
581 /** opaq argument passed to completion \a set_interpret callback. */
584 * Lock for \a set_new_requests manipulations
585 * locked so that any old caller can communicate requests to
586 * the set holder who can then fold them into the lock-free set
588 spinlock_t set_new_req_lock;
589 /** List of new yet unsent requests. Only used with ptlrpcd now. */
590 struct list_head set_new_requests;
592 /** rq_status of requests that have been freed already */
594 /** Additional fields used by the flow control extension */
595 /** Maximum number of RPCs in flight */
596 int set_max_inflight;
597 /** Callback function used to generate RPCs */
598 set_producer_func set_producer;
599 /** opaq argument passed to the producer callback */
600 void *set_producer_arg;
604 * Description of a single ptrlrpc_set callback
606 struct ptlrpc_set_cbdata {
607 /** List linkage item */
608 struct list_head psc_item;
609 /** Pointer to interpreting function */
610 set_interpreter_func psc_interpret;
611 /** Opaq argument to pass to the callback */
615 struct ptlrpc_bulk_desc;
616 struct ptlrpc_service_part;
617 struct ptlrpc_service;
620 * ptlrpc callback & work item stuff
622 struct ptlrpc_cb_id {
623 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
624 void *cbid_arg; /* additional arg */
627 /** Maximum number of locks to fit into reply state */
628 #define RS_MAX_LOCKS 8
632 * Structure to define reply state on the server
633 * Reply state holds various reply message information. Also for "difficult"
634 * replies (rep-ack case) we store the state after sending reply and wait
635 * for the client to acknowledge the reception. In these cases locks could be
636 * added to the state for replay/failover consistency guarantees.
638 struct ptlrpc_reply_state {
639 /** Callback description */
640 struct ptlrpc_cb_id rs_cb_id;
641 /** Linkage for list of all reply states in a system */
642 struct list_head rs_list;
643 /** Linkage for list of all reply states on same export */
644 struct list_head rs_exp_list;
645 /** Linkage for list of all reply states for same obd */
646 struct list_head rs_obd_list;
648 struct list_head rs_debug_list;
650 /** A spinlock to protect the reply state flags */
652 /** Reply state flags */
653 unsigned long rs_difficult:1; /* ACK/commit stuff */
654 unsigned long rs_no_ack:1; /* no ACK, even for
655 difficult requests */
656 unsigned long rs_scheduled:1; /* being handled? */
657 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
658 unsigned long rs_handled:1; /* been handled yet? */
659 unsigned long rs_on_net:1; /* reply_out_callback pending? */
660 unsigned long rs_prealloc:1; /* rs from prealloc list */
661 unsigned long rs_committed:1;/* the transaction was committed
662 and the rs was dispatched
663 by ptlrpc_commit_replies */
664 /** Size of the state */
668 /** Transaction number */
672 struct obd_export *rs_export;
673 struct ptlrpc_service_part *rs_svcpt;
674 /** Lnet metadata handle for the reply */
675 lnet_handle_md_t rs_md_h;
676 atomic_t rs_refcount;
678 /** Context for the sevice thread */
679 struct ptlrpc_svc_ctx *rs_svc_ctx;
680 /** Reply buffer (actually sent to the client), encoded if needed */
681 struct lustre_msg *rs_repbuf; /* wrapper */
682 /** Size of the reply buffer */
683 int rs_repbuf_len; /* wrapper buf length */
684 /** Size of the reply message */
685 int rs_repdata_len; /* wrapper msg length */
687 * Actual reply message. Its content is encrupted (if needed) to
688 * produce reply buffer for actual sending. In simple case
689 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
691 struct lustre_msg *rs_msg; /* reply message */
693 /** Number of locks awaiting client ACK */
695 /** Handles of locks awaiting client reply ACK */
696 struct lustre_handle rs_locks[RS_MAX_LOCKS];
697 /** Lock modes of locks in \a rs_locks */
698 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
701 struct ptlrpc_thread;
705 RQ_PHASE_NEW = 0xebc0de00,
706 RQ_PHASE_RPC = 0xebc0de01,
707 RQ_PHASE_BULK = 0xebc0de02,
708 RQ_PHASE_INTERPRET = 0xebc0de03,
709 RQ_PHASE_COMPLETE = 0xebc0de04,
710 RQ_PHASE_UNREGISTERING = 0xebc0de05,
711 RQ_PHASE_UNDEFINED = 0xebc0de06
714 /** Type of request interpreter call-back */
715 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
716 struct ptlrpc_request *req,
718 /** Type of request resend call-back */
719 typedef void (*ptlrpc_resend_cb_t)(struct ptlrpc_request *req,
723 * Definition of request pool structure.
724 * The pool is used to store empty preallocated requests for the case
725 * when we would actually need to send something without performing
726 * any allocations (to avoid e.g. OOM).
728 struct ptlrpc_request_pool {
729 /** Locks the list */
731 /** list of ptlrpc_request structs */
732 struct list_head prp_req_list;
733 /** Maximum message size that would fit into a rquest from this pool */
735 /** Function to allocate more requests for this pool */
736 void (*prp_populate)(struct ptlrpc_request_pool *, int);
745 * \defgroup nrs Network Request Scheduler
748 struct ptlrpc_nrs_policy;
749 struct ptlrpc_nrs_resource;
750 struct ptlrpc_nrs_request;
753 * NRS control operations.
755 * These are common for all policies.
757 enum ptlrpc_nrs_ctl {
759 * Not a valid opcode.
761 PTLRPC_NRS_CTL_INVALID,
763 * Activate the policy.
765 PTLRPC_NRS_CTL_START,
767 * Reserved for multiple primary policies, which may be a possibility
772 * Policies can start using opcodes from this value and onwards for
773 * their own purposes; the assigned value itself is arbitrary.
775 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
779 * ORR policy operations
782 NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
783 NRS_CTL_ORR_WR_QUANTUM,
784 NRS_CTL_ORR_RD_OFF_TYPE,
785 NRS_CTL_ORR_WR_OFF_TYPE,
786 NRS_CTL_ORR_RD_SUPP_REQ,
787 NRS_CTL_ORR_WR_SUPP_REQ,
791 * NRS policy operations.
793 * These determine the behaviour of a policy, and are called in response to
796 struct ptlrpc_nrs_pol_ops {
798 * Called during policy registration; this operation is optional.
800 * \param[in,out] policy The policy being initialized
802 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
804 * Called during policy unregistration; this operation is optional.
806 * \param[in,out] policy The policy being unregistered/finalized
808 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
810 * Called when activating a policy via lprocfs; policies allocate and
811 * initialize their resources here; this operation is optional.
813 * \param[in,out] policy The policy being started
814 * \param[in,out] arg A generic char buffer
816 * \see nrs_policy_start_locked()
818 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy,
821 * Called when deactivating a policy via lprocfs; policies deallocate
822 * their resources here; this operation is optional
824 * \param[in,out] policy The policy being stopped
826 * \see nrs_policy_stop0()
828 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
830 * Used for policy-specific operations; i.e. not generic ones like
831 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
832 * to an ioctl; this operation is optional.
834 * \param[in,out] policy The policy carrying out operation \a opc
835 * \param[in] opc The command operation being carried out
836 * \param[in,out] arg An generic buffer for communication between the
837 * user and the control operation
842 * \see ptlrpc_nrs_policy_control()
844 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
845 enum ptlrpc_nrs_ctl opc, void *arg);
848 * Called when obtaining references to the resources of the resource
849 * hierarchy for a request that has arrived for handling at the PTLRPC
850 * service. Policies should return -ve for requests they do not wish
851 * to handle. This operation is mandatory.
853 * \param[in,out] policy The policy we're getting resources for.
854 * \param[in,out] nrq The request we are getting resources for.
855 * \param[in] parent The parent resource of the resource being
856 * requested; set to NULL if none.
857 * \param[out] resp The resource is to be returned here; the
858 * fallback policy in an NRS head should
859 * \e always return a non-NULL pointer value.
860 * \param[in] moving_req When set, signifies that this is an attempt
861 * to obtain resources for a request being moved
862 * to the high-priority NRS head by
863 * ldlm_lock_reorder_req().
864 * This implies two things:
865 * 1. We are under obd_export::exp_rpc_lock and
866 * so should not sleep.
867 * 2. We should not perform non-idempotent or can
868 * skip performing idempotent operations that
869 * were carried out when resources were first
870 * taken for the request when it was initialized
871 * in ptlrpc_nrs_req_initialize().
873 * \retval 0, +ve The level of the returned resource in the resource
874 * hierarchy; currently only 0 (for a non-leaf resource)
875 * and 1 (for a leaf resource) are supported by the
879 * \see ptlrpc_nrs_req_initialize()
880 * \see ptlrpc_nrs_hpreq_add_nolock()
881 * \see ptlrpc_nrs_req_hp_move()
883 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
884 struct ptlrpc_nrs_request *nrq,
885 const struct ptlrpc_nrs_resource *parent,
886 struct ptlrpc_nrs_resource **resp,
889 * Called when releasing references taken for resources in the resource
890 * hierarchy for the request; this operation is optional.
892 * \param[in,out] policy The policy the resource belongs to
893 * \param[in] res The resource to be freed
895 * \see ptlrpc_nrs_req_finalize()
896 * \see ptlrpc_nrs_hpreq_add_nolock()
897 * \see ptlrpc_nrs_req_hp_move()
899 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
900 const struct ptlrpc_nrs_resource *res);
903 * Obtains a request for handling from the policy, and optionally
904 * removes the request from the policy; this operation is mandatory.
906 * \param[in,out] policy The policy to poll
907 * \param[in] peek When set, signifies that we just want to
908 * examine the request, and not handle it, so the
909 * request is not removed from the policy.
910 * \param[in] force When set, it will force a policy to return a
911 * request if it has one queued.
913 * \retval NULL No request available for handling
914 * \retval valid-pointer The request polled for handling
916 * \see ptlrpc_nrs_req_get_nolock()
918 struct ptlrpc_nrs_request *
919 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
922 * Called when attempting to add a request to a policy for later
923 * handling; this operation is mandatory.
925 * \param[in,out] policy The policy on which to enqueue \a nrq
926 * \param[in,out] nrq The request to enqueue
931 * \see ptlrpc_nrs_req_add_nolock()
933 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
934 struct ptlrpc_nrs_request *nrq);
936 * Removes a request from the policy's set of pending requests. Normally
937 * called after a request has been polled successfully from the policy
938 * for handling; this operation is mandatory.
940 * \param[in,out] policy The policy the request \a nrq belongs to
941 * \param[in,out] nrq The request to dequeue
943 * \see ptlrpc_nrs_req_del_nolock()
945 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
946 struct ptlrpc_nrs_request *nrq);
948 * Called after the request being carried out. Could be used for
949 * job/resource control; this operation is optional.
951 * \param[in,out] policy The policy which is stopping to handle request
953 * \param[in,out] nrq The request
955 * \pre assert_spin_locked(&svcpt->scp_req_lock)
957 * \see ptlrpc_nrs_req_stop_nolock()
959 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
960 struct ptlrpc_nrs_request *nrq);
962 * Registers the policy's lprocfs interface with a PTLRPC service.
964 * \param[in] svc The service
969 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
971 * Unegisters the policy's lprocfs interface with a PTLRPC service.
973 * In cases of failed policy registration in
974 * \e ptlrpc_nrs_policy_register(), this function may be called for a
975 * service which has not registered the policy successfully, so
976 * implementations of this method should make sure their operations are
977 * safe in such cases.
979 * \param[in] svc The service
981 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
987 enum nrs_policy_flags {
989 * Fallback policy, use this flag only on a single supported policy per
990 * service. The flag cannot be used on policies that use
991 * \e PTLRPC_NRS_FL_REG_EXTERN
993 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
995 * Start policy immediately after registering.
997 PTLRPC_NRS_FL_REG_START = (1 << 1),
999 * This is a policy registering from a module different to the one NRS
1000 * core ships in (currently ptlrpc).
1002 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1008 * Denotes whether an NRS instance is for handling normal or high-priority
1009 * RPCs, or whether an operation pertains to one or both of the NRS instances
1012 enum ptlrpc_nrs_queue_type {
1013 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1014 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1015 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1021 * A PTLRPC service has at least one NRS head instance for handling normal
1022 * priority RPCs, and may optionally have a second NRS head instance for
1023 * handling high-priority RPCs. Each NRS head maintains a list of available
1024 * policies, of which one and only one policy is acting as the fallback policy,
1025 * and optionally a different policy may be acting as the primary policy. For
1026 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1027 * enqueue the RPC using the primary policy (if any). The fallback policy is
1028 * used in the following cases:
1029 * - when there was no primary policy in the
1030 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1032 * - when the primary policy that was at the
1033 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1034 * RPC was initialized, denoted it did not wish, or for some other reason was
1035 * not able to handle the request, by returning a non-valid NRS resource
1037 * - when the primary policy that was at the
1038 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1039 * RPC was initialized, fails later during the request enqueueing stage.
1041 * \see nrs_resource_get_safe()
1042 * \see nrs_request_enqueue()
1045 spinlock_t nrs_lock;
1046 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1048 * List of registered policies
1050 struct list_head nrs_policy_list;
1052 * List of policies with queued requests. Policies that have any
1053 * outstanding requests are queued here, and this list is queried
1054 * in a round-robin manner from NRS core when obtaining a request
1055 * for handling. This ensures that requests from policies that at some
1056 * point transition away from the
1057 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1059 struct list_head nrs_policy_queued;
1061 * Service partition for this NRS head
1063 struct ptlrpc_service_part *nrs_svcpt;
1065 * Primary policy, which is the preferred policy for handling RPCs
1067 struct ptlrpc_nrs_policy *nrs_policy_primary;
1069 * Fallback policy, which is the backup policy for handling RPCs
1071 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1073 * This NRS head handles either HP or regular requests
1075 enum ptlrpc_nrs_queue_type nrs_queue_type;
1077 * # queued requests from all policies in this NRS head
1079 unsigned long nrs_req_queued;
1081 * # scheduled requests from all policies in this NRS head
1083 unsigned long nrs_req_started;
1085 * # policies on this NRS
1087 unsigned nrs_num_pols;
1089 * This NRS head is in progress of starting a policy
1091 unsigned nrs_policy_starting:1;
1093 * In progress of shutting down the whole NRS head; used during
1096 unsigned nrs_stopping:1;
1098 * NRS policy is throttling reqeust
1100 unsigned nrs_throttling:1;
1103 #define NRS_POL_NAME_MAX 16
1105 struct ptlrpc_nrs_pol_desc;
1108 * Service compatibility predicate; this determines whether a policy is adequate
1109 * for handling RPCs of a particular PTLRPC service.
1111 * XXX:This should give the same result during policy registration and
1112 * unregistration, and for all partitions of a service; so the result should not
1113 * depend on temporal service or other properties, that may influence the
1116 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1117 const struct ptlrpc_nrs_pol_desc *desc);
1119 struct ptlrpc_nrs_pol_conf {
1121 * Human-readable policy name
1123 char nc_name[NRS_POL_NAME_MAX];
1125 * NRS operations for this policy
1127 const struct ptlrpc_nrs_pol_ops *nc_ops;
1129 * Service compatibility predicate
1131 nrs_pol_desc_compat_t nc_compat;
1133 * Set for policies that support a single ptlrpc service, i.e. ones that
1134 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1135 * depicts the name of the single service that such policies are
1138 const char *nc_compat_svc_name;
1140 * Owner module for this policy descriptor; policies registering from a
1141 * different module to the one the NRS framework is held within
1142 * (currently ptlrpc), should set this field to THIS_MODULE.
1144 struct module *nc_owner;
1146 * Policy registration flags; a bitmast of \e nrs_policy_flags
1152 * NRS policy registering descriptor
1154 * Is used to hold a description of a policy that can be passed to NRS core in
1155 * order to register the policy with NRS heads in different PTLRPC services.
1157 struct ptlrpc_nrs_pol_desc {
1159 * Human-readable policy name
1161 char pd_name[NRS_POL_NAME_MAX];
1163 * Link into nrs_core::nrs_policies
1165 struct list_head pd_list;
1167 * NRS operations for this policy
1169 const struct ptlrpc_nrs_pol_ops *pd_ops;
1171 * Service compatibility predicate
1173 nrs_pol_desc_compat_t pd_compat;
1175 * Set for policies that are compatible with only one PTLRPC service.
1177 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1179 const char *pd_compat_svc_name;
1181 * Owner module for this policy descriptor.
1183 * We need to hold a reference to the module whenever we might make use
1184 * of any of the module's contents, i.e.
1185 * - If one or more instances of the policy are at a state where they
1186 * might be handling a request, i.e.
1187 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1188 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1189 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1190 * is taken on the module when
1191 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1192 * becomes 0, so that we hold only one reference to the module maximum
1195 * We do not need to hold a reference to the module, even though we
1196 * might use code and data from the module, in the following cases:
1197 * - During external policy registration, because this should happen in
1198 * the module's init() function, in which case the module is safe from
1199 * removal because a reference is being held on the module by the
1200 * kernel, and iirc kmod (and I guess module-init-tools also) will
1201 * serialize any racing processes properly anyway.
1202 * - During external policy unregistration, because this should happen
1203 * in a module's exit() function, and any attempts to start a policy
1204 * instance would need to take a reference on the module, and this is
1205 * not possible once we have reached the point where the exit()
1206 * handler is called.
1207 * - During service registration and unregistration, as service setup
1208 * and cleanup, and policy registration, unregistration and policy
1209 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1210 * as long as users adhere to the convention of registering policies
1211 * in init() and unregistering them in module exit() functions, there
1212 * should not be a race between these operations.
1213 * - During any policy-specific lprocfs operations, because a reference
1214 * is held by the kernel on a proc entry that has been entered by a
1215 * syscall, so as long as proc entries are removed during unregistration time,
1216 * then unregistration and lprocfs operations will be properly
1219 struct module *pd_owner;
1221 * Bitmask of \e nrs_policy_flags
1225 * # of references on this descriptor
1233 * Policies transition from one state to the other during their lifetime
1235 enum ptlrpc_nrs_pol_state {
1237 * Not a valid policy state.
1239 NRS_POL_STATE_INVALID,
1241 * Policies are at this state either at the start of their life, or
1242 * transition here when the user selects a different policy to act
1243 * as the primary one.
1245 NRS_POL_STATE_STOPPED,
1247 * Policy is progress of stopping
1249 NRS_POL_STATE_STOPPING,
1251 * Policy is in progress of starting
1253 NRS_POL_STATE_STARTING,
1255 * A policy is in this state in two cases:
1256 * - it is the fallback policy, which is always in this state.
1257 * - it has been activated by the user; i.e. it is the primary policy,
1259 NRS_POL_STATE_STARTED,
1263 * NRS policy information
1265 * Used for obtaining information for the status of a policy via lprocfs
1267 struct ptlrpc_nrs_pol_info {
1271 char pi_name[NRS_POL_NAME_MAX];
1273 * Current policy state
1275 enum ptlrpc_nrs_pol_state pi_state;
1277 * # RPCs enqueued for later dispatching by the policy
1281 * # RPCs started for dispatch by the policy
1283 long pi_req_started;
1285 * Is this a fallback policy?
1287 unsigned pi_fallback:1;
1293 * There is one instance of this for each policy in each NRS head of each
1294 * PTLRPC service partition.
1296 struct ptlrpc_nrs_policy {
1298 * Linkage into the NRS head's list of policies,
1299 * ptlrpc_nrs:nrs_policy_list
1301 struct list_head pol_list;
1303 * Linkage into the NRS head's list of policies with enqueued
1304 * requests ptlrpc_nrs:nrs_policy_queued
1306 struct list_head pol_list_queued;
1308 * Current state of this policy
1310 enum ptlrpc_nrs_pol_state pol_state;
1312 * Bitmask of nrs_policy_flags
1316 * # RPCs enqueued for later dispatching by the policy
1318 long pol_req_queued;
1320 * # RPCs started for dispatch by the policy
1322 long pol_req_started;
1324 * Usage Reference count taken on the policy instance
1328 * The NRS head this policy has been created at
1330 struct ptlrpc_nrs *pol_nrs;
1332 * Private policy data; varies by policy type
1336 * Policy descriptor for this policy instance.
1338 struct ptlrpc_nrs_pol_desc *pol_desc;
1344 * Resources are embedded into two types of NRS entities:
1345 * - Inside NRS policies, in the policy's private data in
1346 * ptlrpc_nrs_policy::pol_private
1347 * - In objects that act as prime-level scheduling entities in different NRS
1348 * policies; e.g. on a policy that performs round robin or similar order
1349 * scheduling across client NIDs, there would be one NRS resource per unique
1350 * client NID. On a policy which performs round robin scheduling across
1351 * backend filesystem objects, there would be one resource associated with
1352 * each of the backend filesystem objects partaking in the scheduling
1353 * performed by the policy.
1355 * NRS resources share a parent-child relationship, in which resources embedded
1356 * in policy instances are the parent entities, with all scheduling entities
1357 * a policy schedules across being the children, thus forming a simple resource
1358 * hierarchy. This hierarchy may be extended with one or more levels in the
1359 * future if the ability to have more than one primary policy is added.
1361 * Upon request initialization, references to the then active NRS policies are
1362 * taken and used to later handle the dispatching of the request with one of
1365 * \see nrs_resource_get_safe()
1366 * \see ptlrpc_nrs_req_add()
1368 struct ptlrpc_nrs_resource {
1370 * This NRS resource's parent; is NULL for resources embedded in NRS
1371 * policy instances; i.e. those are top-level ones.
1373 struct ptlrpc_nrs_resource *res_parent;
1375 * The policy associated with this resource.
1377 struct ptlrpc_nrs_policy *res_policy;
1390 * This policy is a logical wrapper around previous, non-NRS functionality.
1391 * It dispatches RPCs in the same order as they arrive from the network. This
1392 * policy is currently used as the fallback policy, and the only enabled policy
1393 * on all NRS heads of all PTLRPC service partitions.
1398 * Private data structure for the FIFO policy
1400 struct nrs_fifo_head {
1402 * Resource object for policy instance.
1404 struct ptlrpc_nrs_resource fh_res;
1406 * List of queued requests.
1408 struct list_head fh_list;
1410 * For debugging purposes.
1415 struct nrs_fifo_req {
1416 struct list_head fr_list;
1425 * CRR-N, Client Round Robin over NIDs
1430 * private data structure for CRR-N NRS
1432 struct nrs_crrn_net {
1433 struct ptlrpc_nrs_resource cn_res;
1434 cfs_binheap_t *cn_binheap;
1435 cfs_hash_t *cn_cli_hash;
1437 * Used when a new scheduling round commences, in order to synchronize
1438 * all clients with the new round number.
1442 * Determines the relevant ordering amongst request batches within a
1447 * Round Robin quantum; the maximum number of RPCs that each request
1448 * batch for each client can have in a scheduling round.
1454 * Object representing a client in CRR-N, as identified by its NID
1456 struct nrs_crrn_client {
1457 struct ptlrpc_nrs_resource cc_res;
1458 struct hlist_node cc_hnode;
1461 * The round number against which this client is currently scheduling
1466 * The sequence number used for requests scheduled by this client during
1467 * the current round number.
1472 * Round Robin quantum; the maximum number of RPCs the client is allowed
1473 * to schedule in a single batch of each round.
1477 * # of pending requests for this client, on all existing rounds
1483 * CRR-N NRS request definition
1485 struct nrs_crrn_req {
1487 * Round number for this request; shared with all other requests in the
1492 * Sequence number for this request; shared with all other requests in
1499 * CRR-N policy operations.
1503 * Read the RR quantum size of a CRR-N policy.
1505 NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
1507 * Write the RR quantum size of a CRR-N policy.
1509 NRS_CTL_CRRN_WR_QUANTUM,
1517 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1522 * Lower and upper byte offsets of a brw RPC
1524 struct nrs_orr_req_range {
1530 * RPC types supported by the ORR/TRR policies
1533 NOS_OST_READ = (1 << 0),
1534 NOS_OST_WRITE = (1 << 1),
1535 NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE),
1537 * Default value for policies.
1539 NOS_DFLT = NOS_OST_READ
1543 * As unique keys for grouping RPCs together, we use the object's OST FID for
1544 * the ORR policy, and the OST index for the TRR policy.
1546 * XXX: We waste some space for TRR policy instances by using a union, but it
1547 * allows to consolidate some of the code between ORR and TRR, and these
1548 * policies will probably eventually merge into one anyway.
1550 struct nrs_orr_key {
1552 /** object FID for ORR */
1553 struct lu_fid ok_fid;
1554 /** OST index for TRR */
1560 * The largest base string for unique hash/slab object names is
1561 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1562 * id number, so this _should_ be more than enough for the maximum number of
1563 * CPTs on any system. If it does happen that this statement is incorrect,
1564 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1565 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1566 * will hopefully not go unnoticed.
1568 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1571 * private data structure for ORR and TRR NRS
1573 struct nrs_orr_data {
1574 struct ptlrpc_nrs_resource od_res;
1575 cfs_binheap_t *od_binheap;
1576 cfs_hash_t *od_obj_hash;
1577 struct kmem_cache *od_cache;
1579 * Used when a new scheduling round commences, in order to synchronize
1580 * all object or OST batches with the new round number.
1584 * Determines the relevant ordering amongst request batches within a
1589 * RPC types that are currently supported.
1591 enum nrs_orr_supp od_supp;
1593 * Round Robin quantum; the maxium number of RPCs that each request
1594 * batch for each object or OST can have in a scheduling round.
1598 * Whether to use physical disk offsets or logical file offsets.
1602 * XXX: We need to provide a persistently allocated string to hold
1603 * unique object names for this policy, since in currently supported
1604 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1605 * to the name string provided. kstrdup() is used in the version of
1606 * kmeme_cache_create() in current Linux mainline, so we may be able to
1607 * remove this in the future.
1609 char od_objname[NRS_ORR_OBJ_NAME_MAX];
1613 * Represents a backend-fs object or OST in the ORR and TRR policies
1616 struct nrs_orr_object {
1617 struct ptlrpc_nrs_resource oo_res;
1618 struct hlist_node oo_hnode;
1620 * The round number against which requests are being scheduled for this
1625 * The sequence number used for requests scheduled for this object or
1626 * OST during the current round number.
1630 * The key of the object or OST for which this structure instance is
1633 struct nrs_orr_key oo_key;
1636 * Round Robin quantum; the maximum number of RPCs that are allowed to
1637 * be scheduled for the object or OST in a single batch of each round.
1641 * # of pending requests for this object or OST, on all existing rounds
1647 * ORR/TRR NRS request definition
1649 struct nrs_orr_req {
1651 * The offset range this request covers
1653 struct nrs_orr_req_range or_range;
1655 * Round number for this request; shared with all other requests in the
1660 * Sequence number for this request; shared with all other requests in
1665 * For debugging purposes.
1667 struct nrs_orr_key or_key;
1669 * An ORR policy instance has filled in request information while
1670 * enqueueing the request on the service partition's regular NRS head.
1672 unsigned int or_orr_set:1;
1674 * A TRR policy instance has filled in request information while
1675 * enqueueing the request on the service partition's regular NRS head.
1677 unsigned int or_trr_set:1;
1679 * Request offset ranges have been filled in with logical offset
1682 unsigned int or_logical_set:1;
1684 * Request offset ranges have been filled in with physical offset
1687 unsigned int or_physical_set:1;
1692 #include <lustre_nrs_tbf.h>
1697 * Instances of this object exist embedded within ptlrpc_request; the main
1698 * purpose of this object is to hold references to the request's resources
1699 * for the lifetime of the request, and to hold properties that policies use
1700 * use for determining the request's scheduling priority.
1702 struct ptlrpc_nrs_request {
1704 * The request's resource hierarchy.
1706 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1708 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1709 * policy that was used to enqueue the request.
1711 * \see nrs_request_enqueue()
1713 unsigned nr_res_idx;
1714 unsigned nr_initialized:1;
1715 unsigned nr_enqueued:1;
1716 unsigned nr_started:1;
1717 unsigned nr_finalized:1;
1718 cfs_binheap_node_t nr_node;
1721 * Policy-specific fields, used for determining a request's scheduling
1722 * priority, and other supporting functionality.
1726 * Fields for the FIFO policy
1728 struct nrs_fifo_req fifo;
1730 * CRR-N request defintion
1732 struct nrs_crrn_req crr;
1733 /** ORR and TRR share the same request definition */
1734 struct nrs_orr_req orr;
1736 * TBF request definition
1738 struct nrs_tbf_req tbf;
1741 * Externally-registering policies may want to use this to allocate
1742 * their own request properties.
1750 * Basic request prioritization operations structure.
1751 * The whole idea is centered around locks and RPCs that might affect locks.
1752 * When a lock is contended we try to give priority to RPCs that might lead
1753 * to fastest release of that lock.
1754 * Currently only implemented for OSTs only in a way that makes all
1755 * IO and truncate RPCs that are coming from a locked region where a lock is
1756 * contended a priority over other requests.
1758 struct ptlrpc_hpreq_ops {
1760 * Check if the lock handle of the given lock is the same as
1761 * taken from the request.
1763 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1765 * Check if the request is a high priority one.
1767 int (*hpreq_check)(struct ptlrpc_request *);
1769 * Called after the request has been handled.
1771 void (*hpreq_fini)(struct ptlrpc_request *);
1775 * Represents remote procedure call.
1777 * This is a staple structure used by everybody wanting to send a request
1780 struct ptlrpc_request {
1781 /* Request type: one of PTL_RPC_MSG_* */
1783 /** Result of request processing */
1786 * Linkage item through which this request is included into
1787 * sending/delayed lists on client and into rqbd list on server
1789 struct list_head rq_list;
1791 * Server side list of incoming unserved requests sorted by arrival
1792 * time. Traversed from time to time to notice about to expire
1793 * requests and sent back "early replies" to clients to let them
1794 * know server is alive and well, just very busy to service their
1797 struct list_head rq_timed_list;
1798 /** server-side history, used for debuging purposes. */
1799 struct list_head rq_history_list;
1800 /** server-side per-export list */
1801 struct list_head rq_exp_list;
1802 /** server-side hp handlers */
1803 struct ptlrpc_hpreq_ops *rq_ops;
1805 /** initial thread servicing this request */
1806 struct ptlrpc_thread *rq_svc_thread;
1808 /** history sequence # */
1809 __u64 rq_history_seq;
1813 /** stub for NRS request */
1814 struct ptlrpc_nrs_request rq_nrq;
1816 /** the index of service's srv_at_array into which request is linked */
1818 /** Lock to protect request flags and some other important bits, like
1822 /** client-side flags are serialized by rq_lock */
1823 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1824 rq_timedout:1, rq_resend:1, rq_restart:1,
1826 * when ->rq_replay is set, request is kept by the client even
1827 * after server commits corresponding transaction. This is
1828 * used for operations that require sequence of multiple
1829 * requests to be replayed. The only example currently is file
1830 * open/close. When last request in such a sequence is
1831 * committed, ->rq_replay is cleared on all requests in the
1835 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1836 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1838 rq_req_unlink:1, rq_reply_unlink:1,
1839 rq_memalloc:1, /* req originated from "kswapd" */
1840 /* server-side flags */
1841 rq_packed_final:1, /* packed final reply */
1842 rq_hp:1, /* high priority RPC */
1843 rq_at_linked:1, /* link into service's srv_at_array */
1844 rq_reply_truncate:1,
1846 /* whether the "rq_set" is a valid one */
1848 rq_generation_set:1,
1849 /* do not resend request on -EINPROGRESS */
1850 rq_no_retry_einprogress:1,
1851 /* allow the req to be sent if the import is in recovery
1854 /* bulk request, sent to server, but uncommitted */
1857 unsigned int rq_nr_resend;
1859 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1860 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1861 atomic_t rq_refcount;/* client-side refcount for SENT race,
1862 server-side refcounf for multiple replies */
1864 /** Portal to which this request would be sent */
1865 short rq_request_portal; /* XXX FIXME bug 249 */
1866 /** Portal where to wait for reply and where reply would be sent */
1867 short rq_reply_portal; /* XXX FIXME bug 249 */
1871 * !rq_truncate : # reply bytes actually received,
1872 * rq_truncate : required repbuf_len for resend
1874 int rq_nob_received;
1875 /** Request length */
1879 /** Request message - what client sent */
1880 struct lustre_msg *rq_reqmsg;
1881 /** Reply message - server response */
1882 struct lustre_msg *rq_repmsg;
1883 /** Transaction number */
1888 * List item to for replay list. Not yet commited requests get linked
1890 * Also see \a rq_replay comment above.
1892 struct list_head rq_replay_list;
1895 * security and encryption data
1897 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1898 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1899 struct list_head rq_ctx_chain; /**< link to waited ctx */
1901 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1902 enum lustre_sec_part rq_sp_from;
1904 /* client/server security flags */
1906 rq_ctx_init:1, /* context initiation */
1907 rq_ctx_fini:1, /* context destroy */
1908 rq_bulk_read:1, /* request bulk read */
1909 rq_bulk_write:1, /* request bulk write */
1910 /* server authentication flags */
1911 rq_auth_gss:1, /* authenticated by gss */
1912 rq_auth_remote:1, /* authed as remote user */
1913 rq_auth_usr_root:1, /* authed as root */
1914 rq_auth_usr_mdt:1, /* authed as mdt */
1915 rq_auth_usr_ost:1, /* authed as ost */
1916 /* security tfm flags */
1919 /* doesn't expect reply FIXME */
1921 rq_pill_init:1; /* pill initialized */
1923 uid_t rq_auth_uid; /* authed uid */
1924 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1926 /* (server side), pointed directly into req buffer */
1927 struct ptlrpc_user_desc *rq_user_desc;
1929 /* various buffer pointers */
1930 struct lustre_msg *rq_reqbuf; /* req wrapper */
1931 char *rq_repbuf; /* rep buffer */
1932 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1933 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1934 int rq_reqbuf_len; /* req wrapper buf len */
1935 int rq_reqdata_len; /* req wrapper msg len */
1936 int rq_repbuf_len; /* rep buffer len */
1937 int rq_repdata_len; /* rep wrapper msg len */
1938 int rq_clrbuf_len; /* only in priv mode */
1939 int rq_clrdata_len; /* only in priv mode */
1941 /** early replies go to offset 0, regular replies go after that */
1942 unsigned int rq_reply_off;
1946 /** Fields that help to see if request and reply were swabbed or not */
1947 __u32 rq_req_swab_mask;
1948 __u32 rq_rep_swab_mask;
1950 /** What was import generation when this request was sent */
1951 int rq_import_generation;
1952 enum lustre_imp_state rq_send_state;
1954 /** how many early replies (for stats) */
1957 /** client+server request */
1958 lnet_handle_md_t rq_req_md_h;
1959 struct ptlrpc_cb_id rq_req_cbid;
1960 /** optional time limit for send attempts */
1961 cfs_duration_t rq_delay_limit;
1962 /** time request was first queued */
1963 cfs_time_t rq_queued_time;
1965 /* server-side... */
1966 /** request arrival time */
1967 struct timeval rq_arrival_time;
1968 /** separated reply state */
1969 struct ptlrpc_reply_state *rq_reply_state;
1970 /** incoming request buffer */
1971 struct ptlrpc_request_buffer_desc *rq_rqbd;
1973 /** client-only incoming reply */
1974 lnet_handle_md_t rq_reply_md_h;
1975 wait_queue_head_t rq_reply_waitq;
1976 struct ptlrpc_cb_id rq_reply_cbid;
1980 /** Peer description (the other side) */
1981 lnet_process_id_t rq_peer;
1982 /** Server-side, export on which request was received */
1983 struct obd_export *rq_export;
1984 /** Client side, import where request is being sent */
1985 struct obd_import *rq_import;
1987 /** Replay callback, called after request is replayed at recovery */
1988 void (*rq_replay_cb)(struct ptlrpc_request *);
1990 * Commit callback, called when request is committed and about to be
1993 void (*rq_commit_cb)(struct ptlrpc_request *);
1994 /** Opaq data for replay and commit callbacks. */
1997 /** For bulk requests on client only: bulk descriptor */
1998 struct ptlrpc_bulk_desc *rq_bulk;
2000 /** client outgoing req */
2002 * when request/reply sent (secs), or time when request should be sent
2005 /** time for request really sent out */
2006 time_t rq_real_sent;
2008 /** when request must finish. volatile
2009 * so that servers' early reply updates to the deadline aren't
2010 * kept in per-cpu cache */
2011 volatile time_t rq_deadline;
2012 /** when req reply unlink must finish. */
2013 time_t rq_reply_deadline;
2014 /** when req bulk unlink must finish. */
2015 time_t rq_bulk_deadline;
2017 * service time estimate (secs)
2018 * If the requestsis not served by this time, it is marked as timed out.
2022 /** Multi-rpc bits */
2023 /** Per-request waitq introduced by bug 21938 for recovery waiting */
2024 wait_queue_head_t rq_set_waitq;
2025 /** Link item for request set lists */
2026 struct list_head rq_set_chain;
2027 /** Link back to the request set */
2028 struct ptlrpc_request_set *rq_set;
2029 /** Async completion handler, called when reply is received */
2030 ptlrpc_interpterer_t rq_interpret_reply;
2031 /** Resend handler, called when request is resend to update RPC data */
2032 ptlrpc_resend_cb_t rq_resend_cb;
2033 /** Async completion context */
2034 union ptlrpc_async_args rq_async_args;
2036 /** Pool if request is from preallocated list */
2037 struct ptlrpc_request_pool *rq_pool;
2039 struct lu_context rq_session;
2041 /** request format description */
2042 struct req_capsule rq_pill;
2046 * Call completion handler for rpc if any, return it's status or original
2047 * rc if there was no handler defined for this request.
2049 static inline int ptlrpc_req_interpret(const struct lu_env *env,
2050 struct ptlrpc_request *req, int rc)
2052 if (req->rq_interpret_reply != NULL) {
2053 req->rq_status = req->rq_interpret_reply(env, req,
2054 &req->rq_async_args,
2056 return req->rq_status;
2064 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
2065 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
2066 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
2067 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
2068 struct ptlrpc_nrs_pol_info *info);
2071 * Can the request be moved from the regular NRS head to the high-priority NRS
2072 * head (of the same PTLRPC service partition), if any?
2074 * For a reliable result, this should be checked under svcpt->scp_req lock.
2076 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
2078 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
2081 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
2082 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
2083 * to make sure it has not been scheduled yet (analogous to previous
2084 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
2086 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
2091 * Returns 1 if request buffer at offset \a index was already swabbed
2093 static inline int lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
2095 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2096 return req->rq_req_swab_mask & (1 << index);
2100 * Returns 1 if request reply buffer at offset \a index was already swabbed
2102 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
2104 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2105 return req->rq_rep_swab_mask & (1 << index);
2109 * Returns 1 if request needs to be swabbed into local cpu byteorder
2111 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
2113 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2117 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
2119 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
2121 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2125 * Mark request buffer at offset \a index that it was already swabbed
2127 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
2130 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2131 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
2132 req->rq_req_swab_mask |= 1 << index;
2136 * Mark request reply buffer at offset \a index that it was already swabbed
2138 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
2141 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2142 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
2143 req->rq_rep_swab_mask |= 1 << index;
2147 * Convert numerical request phase value \a phase into text string description
2149 static inline const char *
2150 ptlrpc_phase2str(enum rq_phase phase)
2159 case RQ_PHASE_INTERPRET:
2161 case RQ_PHASE_COMPLETE:
2163 case RQ_PHASE_UNREGISTERING:
2164 return "Unregistering";
2171 * Convert numerical request phase of the request \a req into text stringi
2174 static inline const char *
2175 ptlrpc_rqphase2str(struct ptlrpc_request *req)
2177 return ptlrpc_phase2str(req->rq_phase);
2181 * Debugging functions and helpers to print request structure into debug log
2184 /* Spare the preprocessor, spoil the bugs. */
2185 #define FLAG(field, str) (field ? str : "")
2187 /** Convert bit flags into a string */
2188 #define DEBUG_REQ_FLAGS(req) \
2189 ptlrpc_rqphase2str(req), \
2190 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
2191 FLAG(req->rq_err, "E"), \
2192 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
2193 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
2194 FLAG(req->rq_no_resend, "N"), \
2195 FLAG(req->rq_waiting, "W"), \
2196 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
2197 FLAG(req->rq_committed, "M")
2199 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
2201 void _debug_req(struct ptlrpc_request *req,
2202 struct libcfs_debug_msg_data *data, const char *fmt, ...)
2203 __attribute__ ((format (printf, 3, 4)));
2206 * Helper that decides if we need to print request accordig to current debug
2209 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
2211 CFS_CHECK_STACK(msgdata, mask, cdls); \
2213 if (((mask) & D_CANTMASK) != 0 || \
2214 ((libcfs_debug & (mask)) != 0 && \
2215 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
2216 _debug_req((req), msgdata, fmt, ##a); \
2220 * This is the debug print function you need to use to print request sturucture
2221 * content into lustre debug log.
2222 * for most callers (level is a constant) this is resolved at compile time */
2223 #define DEBUG_REQ(level, req, fmt, args...) \
2225 if ((level) & (D_ERROR | D_WARNING)) { \
2226 static cfs_debug_limit_state_t cdls; \
2227 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
2228 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
2230 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
2231 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
2237 * Structure that defines a single page of a bulk transfer
2239 struct ptlrpc_bulk_page {
2240 /** Linkage to list of pages in a bulk */
2241 struct list_head bp_link;
2243 * Number of bytes in a page to transfer starting from \a bp_pageoffset
2246 /** offset within a page */
2248 /** The page itself */
2249 struct page *bp_page;
2252 #define BULK_GET_SOURCE 0
2253 #define BULK_PUT_SINK 1
2254 #define BULK_GET_SINK 2
2255 #define BULK_PUT_SOURCE 3
2258 * Definition of bulk descriptor.
2259 * Bulks are special "Two phase" RPCs where initial request message
2260 * is sent first and it is followed bt a transfer (o receiving) of a large
2261 * amount of data to be settled into pages referenced from the bulk descriptors.
2262 * Bulks transfers (the actual data following the small requests) are done
2263 * on separate LNet portals.
2264 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2265 * Another user is readpage for MDT.
2267 struct ptlrpc_bulk_desc {
2268 /** completed with failure */
2269 unsigned long bd_failure:1;
2270 /** {put,get}{source,sink} */
2271 unsigned long bd_type:2;
2273 unsigned long bd_registered:1;
2274 /** For serialization with callback */
2276 /** Import generation when request for this bulk was sent */
2277 int bd_import_generation;
2278 /** LNet portal for this bulk */
2280 /** Server side - export this bulk created for */
2281 struct obd_export *bd_export;
2282 /** Client side - import this bulk was sent on */
2283 struct obd_import *bd_import;
2284 /** Back pointer to the request */
2285 struct ptlrpc_request *bd_req;
2286 wait_queue_head_t bd_waitq; /* server side only WQ */
2287 int bd_iov_count; /* # entries in bd_iov */
2288 int bd_max_iov; /* allocated size of bd_iov */
2289 int bd_nob; /* # bytes covered */
2290 int bd_nob_transferred; /* # bytes GOT/PUT */
2294 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2295 lnet_nid_t bd_sender; /* stash event::sender */
2296 int bd_md_count; /* # valid entries in bd_mds */
2297 int bd_md_max_brw; /* max entries in bd_mds */
2298 /** array of associated MDs */
2299 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2302 * encrypt iov, size is either 0 or bd_iov_count.
2304 lnet_kiov_t *bd_enc_iov;
2306 lnet_kiov_t bd_iov[0];
2310 SVC_STOPPED = 1 << 0,
2311 SVC_STOPPING = 1 << 1,
2312 SVC_STARTING = 1 << 2,
2313 SVC_RUNNING = 1 << 3,
2315 SVC_SIGNAL = 1 << 5,
2318 #define PTLRPC_THR_NAME_LEN 32
2320 * Definition of server service thread structure
2322 struct ptlrpc_thread {
2324 * List of active threads in svc->srv_threads
2326 struct list_head t_link;
2328 * thread-private data (preallocated memory)
2333 * service thread index, from ptlrpc_start_threads
2337 * service thread pid
2341 * put watchdog in the structure per thread b=14840
2343 struct lc_watchdog *t_watchdog;
2345 * the svc this thread belonged to b=18582
2347 struct ptlrpc_service_part *t_svcpt;
2348 wait_queue_head_t t_ctl_waitq;
2349 struct lu_env *t_env;
2350 char t_name[PTLRPC_THR_NAME_LEN];
2353 static inline int thread_is_init(struct ptlrpc_thread *thread)
2355 return thread->t_flags == 0;
2358 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2360 return !!(thread->t_flags & SVC_STOPPED);
2363 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2365 return !!(thread->t_flags & SVC_STOPPING);
2368 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2370 return !!(thread->t_flags & SVC_STARTING);
2373 static inline int thread_is_running(struct ptlrpc_thread *thread)
2375 return !!(thread->t_flags & SVC_RUNNING);
2378 static inline int thread_is_event(struct ptlrpc_thread *thread)
2380 return !!(thread->t_flags & SVC_EVENT);
2383 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2385 return !!(thread->t_flags & SVC_SIGNAL);
2388 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2390 thread->t_flags &= ~flags;
2393 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2395 thread->t_flags = flags;
2398 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2400 thread->t_flags |= flags;
2403 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2406 if (thread->t_flags & flags) {
2407 thread->t_flags &= ~flags;
2414 * Request buffer descriptor structure.
2415 * This is a structure that contains one posted request buffer for service.
2416 * Once data land into a buffer, event callback creates actual request and
2417 * notifies wakes one of the service threads to process new incoming request.
2418 * More than one request can fit into the buffer.
2420 struct ptlrpc_request_buffer_desc {
2421 /** Link item for rqbds on a service */
2422 struct list_head rqbd_list;
2423 /** History of requests for this buffer */
2424 struct list_head rqbd_reqs;
2425 /** Back pointer to service for which this buffer is registered */
2426 struct ptlrpc_service_part *rqbd_svcpt;
2427 /** LNet descriptor */
2428 lnet_handle_md_t rqbd_md_h;
2430 /** The buffer itself */
2432 struct ptlrpc_cb_id rqbd_cbid;
2434 * This "embedded" request structure is only used for the
2435 * last request to fit into the buffer
2437 struct ptlrpc_request rqbd_req;
2440 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2442 struct ptlrpc_service_ops {
2444 * if non-NULL called during thread creation (ptlrpc_start_thread())
2445 * to initialize service specific per-thread state.
2447 int (*so_thr_init)(struct ptlrpc_thread *thr);
2449 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2450 * destruct state created by ->srv_init().
2452 void (*so_thr_done)(struct ptlrpc_thread *thr);
2454 * Handler function for incoming requests for this service
2456 int (*so_req_handler)(struct ptlrpc_request *req);
2458 * function to determine priority of the request, it's called
2459 * on every new request
2461 int (*so_hpreq_handler)(struct ptlrpc_request *);
2463 * service-specific print fn
2465 void (*so_req_printer)(void *, struct ptlrpc_request *);
2468 #ifndef __cfs_cacheline_aligned
2469 /* NB: put it here for reducing patche dependence */
2470 # define __cfs_cacheline_aligned
2474 * How many high priority requests to serve before serving one normal
2477 #define PTLRPC_SVC_HP_RATIO 10
2480 * Definition of PortalRPC service.
2481 * The service is listening on a particular portal (like tcp port)
2482 * and perform actions for a specific server like IO service for OST
2483 * or general metadata service for MDS.
2485 struct ptlrpc_service {
2486 /** serialize /proc operations */
2487 spinlock_t srv_lock;
2488 /** most often accessed fields */
2489 /** chain thru all services */
2490 struct list_head srv_list;
2491 /** service operations table */
2492 struct ptlrpc_service_ops srv_ops;
2493 /** only statically allocated strings here; we don't clean them */
2495 /** only statically allocated strings here; we don't clean them */
2496 char *srv_thread_name;
2497 /** service thread list */
2498 struct list_head srv_threads;
2499 /** threads # should be created for each partition on initializing */
2500 int srv_nthrs_cpt_init;
2501 /** limit of threads number for each partition */
2502 int srv_nthrs_cpt_limit;
2503 /** Root of /proc dir tree for this service */
2504 struct proc_dir_entry *srv_procroot;
2505 /** Pointer to statistic data for this service */
2506 struct lprocfs_stats *srv_stats;
2507 /** # hp per lp reqs to handle */
2508 int srv_hpreq_ratio;
2509 /** biggest request to receive */
2510 int srv_max_req_size;
2511 /** biggest reply to send */
2512 int srv_max_reply_size;
2513 /** size of individual buffers */
2515 /** # buffers to allocate in 1 group */
2516 int srv_nbuf_per_group;
2517 /** Local portal on which to receive requests */
2518 __u32 srv_req_portal;
2519 /** Portal on the client to send replies to */
2520 __u32 srv_rep_portal;
2522 * Tags for lu_context associated with this thread, see struct
2526 /** soft watchdog timeout multiplier */
2527 int srv_watchdog_factor;
2528 /** under unregister_service */
2529 unsigned srv_is_stopping:1;
2531 /** max # request buffers in history per partition */
2532 int srv_hist_nrqbds_cpt_max;
2533 /** number of CPTs this service bound on */
2535 /** CPTs array this service bound on */
2537 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2539 /** CPT table this service is running over */
2540 struct cfs_cpt_table *srv_cptable;
2542 * partition data for ptlrpc service
2544 struct ptlrpc_service_part *srv_parts[0];
2548 * Definition of PortalRPC service partition data.
2549 * Although a service only has one instance of it right now, but we
2550 * will have multiple instances very soon (instance per CPT).
2552 * it has four locks:
2554 * serialize operations on rqbd and requests waiting for preprocess
2556 * serialize operations active requests sent to this portal
2558 * serialize adaptive timeout stuff
2560 * serialize operations on RS list (reply states)
2562 * We don't have any use-case to take two or more locks at the same time
2563 * for now, so there is no lock order issue.
2565 struct ptlrpc_service_part {
2566 /** back reference to owner */
2567 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2568 /* CPT id, reserved */
2570 /** always increasing number */
2572 /** # of starting threads */
2573 int scp_nthrs_starting;
2574 /** # of stopping threads, reserved for shrinking threads */
2575 int scp_nthrs_stopping;
2576 /** # running threads */
2577 int scp_nthrs_running;
2578 /** service threads list */
2579 struct list_head scp_threads;
2582 * serialize the following fields, used for protecting
2583 * rqbd list and incoming requests waiting for preprocess,
2584 * threads starting & stopping are also protected by this lock.
2586 spinlock_t scp_lock __cfs_cacheline_aligned;
2587 /** total # req buffer descs allocated */
2588 int scp_nrqbds_total;
2589 /** # posted request buffers for receiving */
2590 int scp_nrqbds_posted;
2591 /** in progress of allocating rqbd */
2592 int scp_rqbd_allocating;
2593 /** # incoming reqs */
2594 int scp_nreqs_incoming;
2595 /** request buffers to be reposted */
2596 struct list_head scp_rqbd_idle;
2597 /** req buffers receiving */
2598 struct list_head scp_rqbd_posted;
2599 /** incoming reqs */
2600 struct list_head scp_req_incoming;
2601 /** timeout before re-posting reqs, in tick */
2602 cfs_duration_t scp_rqbd_timeout;
2604 * all threads sleep on this. This wait-queue is signalled when new
2605 * incoming request arrives and when difficult reply has to be handled.
2607 wait_queue_head_t scp_waitq;
2609 /** request history */
2610 struct list_head scp_hist_reqs;
2611 /** request buffer history */
2612 struct list_head scp_hist_rqbds;
2613 /** # request buffers in history */
2614 int scp_hist_nrqbds;
2615 /** sequence number for request */
2617 /** highest seq culled from history */
2618 __u64 scp_hist_seq_culled;
2621 * serialize the following fields, used for processing requests
2622 * sent to this portal
2624 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2625 /** # reqs in either of the NRS heads below */
2626 /** # reqs being served */
2627 int scp_nreqs_active;
2628 /** # HPreqs being served */
2629 int scp_nhreqs_active;
2630 /** # hp requests handled */
2633 /** NRS head for regular requests */
2634 struct ptlrpc_nrs scp_nrs_reg;
2635 /** NRS head for HP requests; this is only valid for services that can
2636 * handle HP requests */
2637 struct ptlrpc_nrs *scp_nrs_hp;
2642 * serialize the following fields, used for changes on
2645 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2646 /** estimated rpc service time */
2647 struct adaptive_timeout scp_at_estimate;
2648 /** reqs waiting for replies */
2649 struct ptlrpc_at_array scp_at_array;
2650 /** early reply timer */
2651 struct timer_list scp_at_timer;
2653 cfs_time_t scp_at_checktime;
2654 /** check early replies */
2655 unsigned scp_at_check;
2659 * serialize the following fields, used for processing
2660 * replies for this portal
2662 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2663 /** all the active replies */
2664 struct list_head scp_rep_active;
2665 /** List of free reply_states */
2666 struct list_head scp_rep_idle;
2667 /** waitq to run, when adding stuff to srv_free_rs_list */
2668 wait_queue_head_t scp_rep_waitq;
2669 /** # 'difficult' replies */
2670 atomic_t scp_nreps_difficult;
2673 #define ptlrpc_service_for_each_part(part, i, svc) \
2675 i < (svc)->srv_ncpts && \
2676 (svc)->srv_parts != NULL && \
2677 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2680 * Declaration of ptlrpcd control structure
2682 struct ptlrpcd_ctl {
2684 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2686 unsigned long pc_flags;
2688 * Thread lock protecting structure fields.
2694 struct completion pc_starting;
2698 struct completion pc_finishing;
2700 * Thread requests set.
2702 struct ptlrpc_request_set *pc_set;
2704 * Thread name used in kthread_run()
2708 * Environment for request interpreters to run in.
2710 struct lu_env pc_env;
2712 * Index of ptlrpcd thread in the array.
2716 * Number of the ptlrpcd's partners.
2720 * Pointer to the array of partners' ptlrpcd_ctl structure.
2722 struct ptlrpcd_ctl **pc_partners;
2724 * Record the partner index to be processed next.
2729 /* Bits for pc_flags */
2730 enum ptlrpcd_ctl_flags {
2732 * Ptlrpc thread start flag.
2734 LIOD_START = 1 << 0,
2736 * Ptlrpc thread stop flag.
2740 * Ptlrpc thread force flag (only stop force so far).
2741 * This will cause aborting any inflight rpcs handled
2742 * by thread if LIOD_STOP is specified.
2744 LIOD_FORCE = 1 << 2,
2746 * This is a recovery ptlrpc thread.
2748 LIOD_RECOVERY = 1 << 3,
2750 * The ptlrpcd is bound to some CPU core.
2759 * Service compatibility function; the policy is compatible with all services.
2761 * \param[in] svc The service the policy is attempting to register with.
2762 * \param[in] desc The policy descriptor
2764 * \retval true The policy is compatible with the service
2766 * \see ptlrpc_nrs_pol_desc::pd_compat()
2768 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2769 const struct ptlrpc_nrs_pol_desc *desc)
2775 * Service compatibility function; the policy is compatible with only a specific
2776 * service which is identified by its human-readable name at
2777 * ptlrpc_service::srv_name.
2779 * \param[in] svc The service the policy is attempting to register with.
2780 * \param[in] desc The policy descriptor
2782 * \retval false The policy is not compatible with the service
2783 * \retval true The policy is compatible with the service
2785 * \see ptlrpc_nrs_pol_desc::pd_compat()
2787 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2788 const struct ptlrpc_nrs_pol_desc *desc)
2790 LASSERT(desc->pd_compat_svc_name != NULL);
2791 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2796 /* ptlrpc/events.c */
2797 extern lnet_handle_eq_t ptlrpc_eq_h;
2798 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2799 lnet_process_id_t *peer, lnet_nid_t *self);
2801 * These callbacks are invoked by LNet when something happened to
2805 extern void request_out_callback(lnet_event_t *ev);
2806 extern void reply_in_callback(lnet_event_t *ev);
2807 extern void client_bulk_callback(lnet_event_t *ev);
2808 extern void request_in_callback(lnet_event_t *ev);
2809 extern void reply_out_callback(lnet_event_t *ev);
2810 #ifdef HAVE_SERVER_SUPPORT
2811 extern void server_bulk_callback(lnet_event_t *ev);
2815 /* ptlrpc/connection.c */
2816 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2818 struct obd_uuid *uuid);
2819 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2820 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2821 int ptlrpc_connection_init(void);
2822 void ptlrpc_connection_fini(void);
2823 extern lnet_pid_t ptl_get_pid(void);
2825 /* ptlrpc/niobuf.c */
2827 * Actual interfacing with LNet to put/get/register/unregister stuff
2830 #ifdef HAVE_SERVER_SUPPORT
2831 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2832 unsigned npages, unsigned max_brw,
2833 unsigned type, unsigned portal);
2834 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2835 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2837 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2841 LASSERT(desc != NULL);
2843 spin_lock(&desc->bd_lock);
2844 rc = desc->bd_md_count;
2845 spin_unlock(&desc->bd_lock);
2850 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2851 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2853 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2855 struct ptlrpc_bulk_desc *desc;
2858 LASSERT(req != NULL);
2859 desc = req->rq_bulk;
2861 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2862 req->rq_bulk_deadline > cfs_time_current_sec())
2868 spin_lock(&desc->bd_lock);
2869 rc = desc->bd_md_count;
2870 spin_unlock(&desc->bd_lock);
2874 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2875 #define PTLRPC_REPLY_EARLY 0x02
2876 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2877 int ptlrpc_reply(struct ptlrpc_request *req);
2878 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2879 int ptlrpc_error(struct ptlrpc_request *req);
2880 void ptlrpc_resend_req(struct ptlrpc_request *request);
2881 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2882 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2883 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2886 /* ptlrpc/client.c */
2888 * Client-side portals API. Everything to send requests, receive replies,
2889 * request queues, request management, etc.
2892 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2894 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2895 struct ptlrpc_client *);
2896 void ptlrpc_cleanup_client(struct obd_import *imp);
2897 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2899 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2900 int ptlrpc_replay_req(struct ptlrpc_request *req);
2901 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2902 void ptlrpc_restart_req(struct ptlrpc_request *req);
2903 void ptlrpc_abort_inflight(struct obd_import *imp);
2904 void ptlrpc_cleanup_imp(struct obd_import *imp);
2905 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2907 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2908 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2910 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2911 set_interpreter_func fn, void *data);
2912 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2913 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2914 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2915 int ptlrpc_expired_set(void *data);
2916 void ptlrpc_interrupted_set(void *data);
2917 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2918 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2919 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2920 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2921 struct ptlrpc_request *req);
2923 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2924 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2926 struct ptlrpc_request_pool *
2927 ptlrpc_init_rq_pool(int, int,
2928 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2930 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2931 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2932 const struct req_format *format);
2933 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2934 struct ptlrpc_request_pool *,
2935 const struct req_format *format);
2936 void ptlrpc_request_free(struct ptlrpc_request *request);
2937 int ptlrpc_request_pack(struct ptlrpc_request *request,
2938 __u32 version, int opcode);
2939 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2940 const struct req_format *format,
2941 __u32 version, int opcode);
2942 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2943 __u32 version, int opcode, char **bufs,
2944 struct ptlrpc_cli_ctx *ctx);
2945 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2946 int opcode, int count, __u32 *lengths,
2948 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2949 __u32 version, int opcode,
2950 int count, __u32 *lengths, char **bufs,
2951 struct ptlrpc_request_pool *pool);
2952 void ptlrpc_req_finished(struct ptlrpc_request *request);
2953 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2954 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2955 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2956 unsigned npages, unsigned max_brw,
2957 unsigned type, unsigned portal);
2958 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2959 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2961 __ptlrpc_free_bulk(bulk, 1);
2963 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2965 __ptlrpc_free_bulk(bulk, 0);
2967 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2968 struct page *page, int pageoffset, int len, int);
2969 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2970 struct page *page, int pageoffset,
2973 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2976 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2977 struct page *page, int pageoffset,
2980 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2983 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2984 struct obd_import *imp);
2985 __u64 ptlrpc_next_xid(void);
2986 __u64 ptlrpc_sample_next_xid(void);
2987 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2989 /* Set of routines to run a function in ptlrpcd context */
2990 void *ptlrpcd_alloc_work(struct obd_import *imp,
2991 int (*cb)(const struct lu_env *, void *), void *data);
2992 void ptlrpcd_destroy_work(void *handler);
2993 int ptlrpcd_queue_work(void *handler);
2996 struct ptlrpc_service_buf_conf {
2997 /* nbufs is buffers # to allocate when growing the pool */
2998 unsigned int bc_nbufs;
2999 /* buffer size to post */
3000 unsigned int bc_buf_size;
3001 /* portal to listed for requests on */
3002 unsigned int bc_req_portal;
3003 /* portal of where to send replies to */
3004 unsigned int bc_rep_portal;
3005 /* maximum request size to be accepted for this service */
3006 unsigned int bc_req_max_size;
3007 /* maximum reply size this service can ever send */
3008 unsigned int bc_rep_max_size;
3011 struct ptlrpc_service_thr_conf {
3012 /* threadname should be 8 characters or less - 6 will be added on */
3014 /* threads increasing factor for each CPU */
3015 unsigned int tc_thr_factor;
3016 /* service threads # to start on each partition while initializing */
3017 unsigned int tc_nthrs_init;
3019 * low water of threads # upper-limit on each partition while running,
3020 * service availability may be impacted if threads number is lower
3021 * than this value. It can be ZERO if the service doesn't require
3022 * CPU affinity or there is only one partition.
3024 unsigned int tc_nthrs_base;
3025 /* "soft" limit for total threads number */
3026 unsigned int tc_nthrs_max;
3027 /* user specified threads number, it will be validated due to
3028 * other members of this structure. */
3029 unsigned int tc_nthrs_user;
3030 /* set NUMA node affinity for service threads */
3031 unsigned int tc_cpu_affinity;
3032 /* Tags for lu_context associated with service thread */
3036 struct ptlrpc_service_cpt_conf {
3037 struct cfs_cpt_table *cc_cptable;
3038 /* string pattern to describe CPTs for a service */
3042 struct ptlrpc_service_conf {
3045 /* soft watchdog timeout multiplifier to print stuck service traces */
3046 unsigned int psc_watchdog_factor;
3047 /* buffer information */
3048 struct ptlrpc_service_buf_conf psc_buf;
3049 /* thread information */
3050 struct ptlrpc_service_thr_conf psc_thr;
3051 /* CPU partition information */
3052 struct ptlrpc_service_cpt_conf psc_cpt;
3053 /* function table */
3054 struct ptlrpc_service_ops psc_ops;
3057 /* ptlrpc/service.c */
3059 * Server-side services API. Register/unregister service, request state
3060 * management, service thread management
3064 void ptlrpc_save_lock(struct ptlrpc_request *req,
3065 struct lustre_handle *lock, int mode, int no_ack);
3066 void ptlrpc_commit_replies(struct obd_export *exp);
3067 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
3068 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
3069 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
3070 struct ptlrpc_service *ptlrpc_register_service(
3071 struct ptlrpc_service_conf *conf,
3072 struct proc_dir_entry *proc_entry);
3073 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
3075 int ptlrpc_start_threads(struct ptlrpc_service *svc);
3076 int ptlrpc_unregister_service(struct ptlrpc_service *service);
3077 int liblustre_check_services(void *arg);
3078 void ptlrpc_daemonize(char *name);
3079 int ptlrpc_service_health_check(struct ptlrpc_service *);
3080 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
3081 void ptlrpc_request_change_export(struct ptlrpc_request *req,
3082 struct obd_export *export);
3083 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay);
3085 int ptlrpc_hr_init(void);
3086 void ptlrpc_hr_fini(void);
3090 /* ptlrpc/import.c */
3095 int ptlrpc_connect_import(struct obd_import *imp);
3096 int ptlrpc_init_import(struct obd_import *imp);
3097 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
3098 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
3099 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
3102 /* ptlrpc/pack_generic.c */
3103 int ptlrpc_reconnect_import(struct obd_import *imp);
3107 * ptlrpc msg buffer and swab interface
3111 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
3113 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
3115 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
3116 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
3118 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
3119 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
3121 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
3122 __u32 *lens, char **bufs);
3123 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
3125 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
3126 __u32 *lens, char **bufs, int flags);
3127 #define LPRFL_EARLY_REPLY 1
3128 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
3129 char **bufs, int flags);
3130 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
3131 unsigned int newlen, int move_data);
3132 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
3133 int __lustre_unpack_msg(struct lustre_msg *m, int len);
3134 int lustre_msg_hdr_size(__u32 magic, int count);
3135 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
3136 int lustre_msg_size_v2(int count, __u32 *lengths);
3137 int lustre_packed_msg_size(struct lustre_msg *msg);
3138 int lustre_msg_early_size(void);
3139 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
3140 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
3141 int lustre_msg_buflen(struct lustre_msg *m, int n);
3142 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
3143 int lustre_msg_bufcount(struct lustre_msg *m);
3144 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
3145 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
3146 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
3147 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
3148 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
3149 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
3150 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
3151 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
3152 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
3153 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
3154 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
3155 __u32 lustre_msg_get_type(struct lustre_msg *msg);
3156 __u32 lustre_msg_get_version(struct lustre_msg *msg);
3157 void lustre_msg_add_version(struct lustre_msg *msg, int version);
3158 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
3159 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
3160 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
3161 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
3162 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
3163 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
3164 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
3165 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
3166 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
3167 int lustre_msg_get_status(struct lustre_msg *msg);
3168 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
3169 int lustre_msg_is_v1(struct lustre_msg *msg);
3170 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
3171 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
3172 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
3173 char *lustre_msg_get_jobid(struct lustre_msg *msg);
3174 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
3175 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 53, 0)
3176 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
3178 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
3180 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
3181 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
3182 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
3183 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
3184 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
3185 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
3186 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
3187 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
3188 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
3189 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
3190 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
3191 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
3192 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
3193 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
3194 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
3197 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
3198 unsigned int newlen, int move_data)
3200 LASSERT(req->rq_reply_state);
3201 LASSERT(req->rq_repmsg);
3202 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
3206 #ifdef LUSTRE_TRANSLATE_ERRNOS
3208 static inline int ptlrpc_status_hton(int h)
3211 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
3212 * ELDLM_LOCK_ABORTED, etc.
3215 return -lustre_errno_hton(-h);
3220 static inline int ptlrpc_status_ntoh(int n)
3223 * See the comment in ptlrpc_status_hton().
3226 return -lustre_errno_ntoh(-n);
3233 #define ptlrpc_status_hton(h) (h)
3234 #define ptlrpc_status_ntoh(n) (n)
3239 /** Change request phase of \a req to \a new_phase */
3241 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
3243 if (req->rq_phase == new_phase)
3246 if (new_phase == RQ_PHASE_UNREGISTERING) {
3247 req->rq_next_phase = req->rq_phase;
3249 atomic_inc(&req->rq_import->imp_unregistering);
3252 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
3254 atomic_dec(&req->rq_import->imp_unregistering);
3257 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
3258 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
3260 req->rq_phase = new_phase;
3264 * Returns true if request \a req got early reply and hard deadline is not met
3267 ptlrpc_client_early(struct ptlrpc_request *req)
3269 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3270 req->rq_reply_deadline > cfs_time_current_sec())
3272 return req->rq_early;
3276 * Returns true if we got real reply from server for this request
3279 ptlrpc_client_replied(struct ptlrpc_request *req)
3281 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3282 req->rq_reply_deadline > cfs_time_current_sec())
3284 return req->rq_replied;
3287 /** Returns true if request \a req is in process of receiving server reply */
3289 ptlrpc_client_recv(struct ptlrpc_request *req)
3291 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3292 req->rq_reply_deadline > cfs_time_current_sec())
3294 return req->rq_receiving_reply;
3298 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3302 spin_lock(&req->rq_lock);
3303 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3304 req->rq_reply_deadline > cfs_time_current_sec()) {
3305 spin_unlock(&req->rq_lock);
3308 rc = req->rq_receiving_reply ;
3309 rc = rc || req->rq_req_unlink || req->rq_reply_unlink;
3310 spin_unlock(&req->rq_lock);
3315 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3317 if (req->rq_set == NULL)
3318 wake_up(&req->rq_reply_waitq);
3320 wake_up(&req->rq_set->set_waitq);
3324 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3326 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3327 atomic_inc(&rs->rs_refcount);
3331 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3333 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3334 if (atomic_dec_and_test(&rs->rs_refcount))
3335 lustre_free_reply_state(rs);
3338 /* Should only be called once per req */
3339 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3341 if (req->rq_reply_state == NULL)
3342 return; /* shouldn't occur */
3343 ptlrpc_rs_decref(req->rq_reply_state);
3344 req->rq_reply_state = NULL;
3345 req->rq_repmsg = NULL;
3348 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3350 return lustre_msg_get_magic(req->rq_reqmsg);
3353 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3355 switch (req->rq_reqmsg->lm_magic) {
3356 case LUSTRE_MSG_MAGIC_V2:
3357 return req->rq_reqmsg->lm_repsize;
3359 LASSERTF(0, "incorrect message magic: %08x\n",
3360 req->rq_reqmsg->lm_magic);
3365 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3367 if (req->rq_delay_limit != 0 &&
3368 cfs_time_before(cfs_time_add(req->rq_queued_time,
3369 cfs_time_seconds(req->rq_delay_limit)),
3370 cfs_time_current())) {
3376 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3378 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3379 spin_lock(&req->rq_lock);
3380 req->rq_no_resend = 1;
3381 spin_unlock(&req->rq_lock);
3383 return req->rq_no_resend;
3387 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3389 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3391 return svcpt->scp_service->srv_watchdog_factor *
3392 max_t(int, at, obd_timeout);
3395 static inline struct ptlrpc_service *
3396 ptlrpc_req2svc(struct ptlrpc_request *req)
3398 LASSERT(req->rq_rqbd != NULL);
3399 return req->rq_rqbd->rqbd_svcpt->scp_service;
3402 /* ldlm/ldlm_lib.c */
3404 * Target client logic
3407 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3408 int client_obd_cleanup(struct obd_device *obddev);
3409 int client_connect_import(const struct lu_env *env,
3410 struct obd_export **exp, struct obd_device *obd,
3411 struct obd_uuid *cluuid, struct obd_connect_data *,
3413 int client_disconnect_export(struct obd_export *exp);
3414 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3416 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3417 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3418 struct obd_uuid *uuid);
3419 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3420 void client_destroy_import(struct obd_import *imp);
3423 #ifdef HAVE_SERVER_SUPPORT
3424 int server_disconnect_export(struct obd_export *exp);
3427 /* ptlrpc/pinger.c */
3429 * Pinger API (client side only)
3432 enum timeout_event {
3435 struct timeout_item;
3436 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3437 int ptlrpc_pinger_add_import(struct obd_import *imp);
3438 int ptlrpc_pinger_del_import(struct obd_import *imp);
3439 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3440 timeout_cb_t cb, void *data,
3441 struct list_head *obd_list);
3442 int ptlrpc_del_timeout_client(struct list_head *obd_list,
3443 enum timeout_event event);
3444 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3445 int ptlrpc_obd_ping(struct obd_device *obd);
3446 void ping_evictor_start(void);
3447 void ping_evictor_stop(void);
3448 void ptlrpc_pinger_ir_up(void);
3449 void ptlrpc_pinger_ir_down(void);
3451 int ptlrpc_pinger_suppress_pings(void);
3453 /* ptlrpc daemon bind policy */
3455 /* all ptlrpcd threads are free mode */
3456 PDB_POLICY_NONE = 1,
3457 /* all ptlrpcd threads are bound mode */
3458 PDB_POLICY_FULL = 2,
3459 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3460 PDB_POLICY_PAIR = 3,
3461 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3462 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3463 * If kernel supports NUMA, pthrpcd threads are binded and
3464 * grouped by NUMA node */
3465 PDB_POLICY_NEIGHBOR = 4,
3468 /* ptlrpc daemon load policy
3469 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3470 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3471 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3472 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3473 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3475 /* on the same CPU core as the caller */
3476 PDL_POLICY_SAME = 1,
3477 /* within the same CPU partition, but not the same core as the caller */
3478 PDL_POLICY_LOCAL = 2,
3479 /* round-robin on all CPU cores, but not the same core as the caller */
3480 PDL_POLICY_ROUND = 3,
3481 /* the specified CPU core is preferred, but not enforced */
3482 PDL_POLICY_PREFERRED = 4,
3485 /* ptlrpc/ptlrpcd.c */
3486 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3487 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3488 void ptlrpcd_wake(struct ptlrpc_request *req);
3489 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3490 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3491 int ptlrpcd_addref(void);
3492 void ptlrpcd_decref(void);
3494 /* ptlrpc/lproc_ptlrpc.c */
3496 * procfs output related functions
3499 const char* ll_opcode2str(__u32 opcode);
3501 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3502 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3503 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3505 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3506 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3507 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3511 /* ptlrpc/llog_server.c */
3512 int llog_origin_handle_open(struct ptlrpc_request *req);
3513 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3514 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3515 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3516 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3517 int llog_origin_handle_close(struct ptlrpc_request *req);
3519 /* ptlrpc/llog_client.c */
3520 extern struct llog_operations llog_client_ops;