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, 2012, 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 <linux/lustre_net.h>
60 #include <linux/libcfs/libcfs.h>
62 #include <linux/lnet/lnet.h>
63 #include <lustre/lustre_idl.h>
64 #include <lustre_ha.h>
65 #include <lustre_sec.h>
66 #include <lustre_import.h>
67 #include <lprocfs_status.h>
68 #include <lu_object.h>
69 #include <lustre_req_layout.h>
71 #include <obd_support.h>
72 #include <lustre_ver.h>
74 /* MD flags we _always_ use */
75 #define PTLRPC_MD_OPTIONS 0
78 * Max # of bulk operations in one request.
79 * In order for the client and server to properly negotiate the maximum
80 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
81 * value. The client is free to limit the actual RPC size for any bulk
82 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
83 #define PTLRPC_BULK_OPS_BITS 2
84 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
86 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
87 * should not be used on the server at all. Otherwise, it imposes a
88 * protocol limitation on the maximum RPC size that can be used by any
89 * RPC sent to that server in the future. Instead, the server should
90 * use the negotiated per-client ocd_brw_size to determine the bulk
92 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
95 * Define maxima for bulk I/O.
97 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
98 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
99 * currently supported maximum between peers at connect via ocd_brw_size.
101 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
102 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
103 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
105 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
106 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
107 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
108 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
109 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
110 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
112 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
113 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
114 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
116 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
117 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
119 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_SIZE too big"
122 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
123 # error "PTLRPC_MAX_BRW_PAGES too big"
126 #define PTLRPC_NTHRS_INIT 2
131 * Constants determine how memory is used to buffer incoming service requests.
133 * ?_NBUFS # buffers to allocate when growing the pool
134 * ?_BUFSIZE # bytes in a single request buffer
135 * ?_MAXREQSIZE # maximum request service will receive
137 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
138 * of ?_NBUFS is added to the pool.
140 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
141 * considered full when less than ?_MAXREQSIZE is left in them.
146 * Constants determine how threads are created for ptlrpc service.
148 * ?_NTHRS_INIT # threads to create for each service partition on
149 * initializing. If it's non-affinity service and
150 * there is only one partition, it's the overall #
151 * threads for the service while initializing.
152 * ?_NTHRS_BASE # threads should be created at least for each
153 * ptlrpc partition to keep the service healthy.
154 * It's the low-water mark of threads upper-limit
155 * for each partition.
156 * ?_THR_FACTOR # threads can be added on threads upper-limit for
157 * each CPU core. This factor is only for reference,
158 * we might decrease value of factor if number of cores
159 * per CPT is above a limit.
160 * ?_NTHRS_MAX # overall threads can be created for a service,
161 * it's a soft limit because if service is running
162 * on machine with hundreds of cores and tens of
163 * CPU partitions, we need to guarantee each partition
164 * has ?_NTHRS_BASE threads, which means total threads
165 * will be ?_NTHRS_BASE * number_of_cpts which can
166 * exceed ?_NTHRS_MAX.
170 * #define MDS_NTHRS_INIT 2
171 * #define MDS_NTHRS_BASE 64
172 * #define MDS_NTHRS_FACTOR 8
173 * #define MDS_NTHRS_MAX 1024
176 * ---------------------------------------------------------------------
177 * Server(A) has 16 cores, user configured it to 4 partitions so each
178 * partition has 4 cores, then actual number of service threads on each
180 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
182 * Total number of threads for the service is:
183 * 96 * partitions(4) = 384
186 * ---------------------------------------------------------------------
187 * Server(B) has 32 cores, user configured it to 4 partitions so each
188 * partition has 8 cores, then actual number of service threads on each
190 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
192 * Total number of threads for the service is:
193 * 128 * partitions(4) = 512
196 * ---------------------------------------------------------------------
197 * Server(B) has 96 cores, user configured it to 8 partitions so each
198 * partition has 12 cores, then actual number of service threads on each
200 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
202 * Total number of threads for the service is:
203 * 160 * partitions(8) = 1280
205 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
206 * as upper limit of threads number for each partition:
207 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
210 * ---------------------------------------------------------------------
211 * Server(C) have a thousand of cores and user configured it to 32 partitions
212 * MDS_NTHRS_BASE(64) * 32 = 2048
214 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
215 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
216 * to keep service healthy, so total number of threads will just be 2048.
218 * NB: we don't suggest to choose server with that many cores because backend
219 * filesystem itself, buffer cache, or underlying network stack might
220 * have some SMP scalability issues at that large scale.
222 * If user already has a fat machine with hundreds or thousands of cores,
223 * there are two choices for configuration:
224 * a) create CPU table from subset of all CPUs and run Lustre on
226 * b) bind service threads on a few partitions, see modparameters of
227 * MDS and OSS for details
229 * NB: these calculations (and examples below) are simplified to help
230 * understanding, the real implementation is a little more complex,
231 * please see ptlrpc_server_nthreads_check() for details.
236 * LDLM threads constants:
238 * Given 8 as factor and 24 as base threads number
241 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
244 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
245 * threads for each partition and total threads number will be 112.
248 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
249 * threads for each partition to keep service healthy, so total threads
250 * number should be 24 * 8 = 192.
252 * So with these constants, threads number will be at the similar level
253 * of old versions, unless target machine has over a hundred cores
255 #define LDLM_THR_FACTOR 8
256 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
257 #define LDLM_NTHRS_BASE 24
258 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
260 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
261 #define LDLM_CLIENT_NBUFS 1
262 #define LDLM_SERVER_NBUFS 64
263 #define LDLM_BUFSIZE (8 * 1024)
264 #define LDLM_MAXREQSIZE (5 * 1024)
265 #define LDLM_MAXREPSIZE (1024)
268 * MDS threads constants:
270 * Please see examples in "Thread Constants", MDS threads number will be at
271 * the comparable level of old versions, unless the server has many cores.
273 #ifndef MDS_MAX_THREADS
274 #define MDS_MAX_THREADS 1024
275 #define MDS_MAX_OTHR_THREADS 256
277 #else /* MDS_MAX_THREADS */
278 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
279 #undef MDS_MAX_THREADS
280 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
282 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
285 /* default service */
286 #define MDS_THR_FACTOR 8
287 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
288 #define MDS_NTHRS_MAX MDS_MAX_THREADS
289 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
291 /* read-page service */
292 #define MDS_RDPG_THR_FACTOR 4
293 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
294 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
295 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
297 /* these should be removed when we remove setattr service in the future */
298 #define MDS_SETA_THR_FACTOR 4
299 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
300 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
301 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
303 /* non-affinity threads */
304 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
305 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
310 * Assume file name length = FNAME_MAX = 256 (true for ext3).
311 * path name length = PATH_MAX = 4096
312 * LOV MD size max = EA_MAX = 24 * 2000
313 * (NB: 24 is size of lov_ost_data)
314 * LOV LOGCOOKIE size max = 32 * 2000
315 * (NB: 32 is size of llog_cookie)
316 * symlink: FNAME_MAX + PATH_MAX <- largest
317 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
318 * rename: FNAME_MAX + FNAME_MAX
319 * open: FNAME_MAX + EA_MAX
321 * MDS_MAXREQSIZE ~= 4736 bytes =
322 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
323 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
325 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
326 * except in the open case where there are a large number of OSTs in a LOV.
328 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
329 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
332 * MDS incoming request with LOV EA
333 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
335 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
336 362 + LOV_MAX_STRIPE_COUNT * 24)
338 * MDS outgoing reply with LOV EA
340 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
341 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
343 * but 2.4 or later MDS will never send reply with llog_cookie to any
344 * version client. This macro is defined for server side reply buffer size.
346 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
349 * This is the size of a maximum REINT_SETXATTR request:
351 * lustre_msg 56 (32 + 4 x 5 + 4)
353 * mdt_rec_setxattr 136
355 * name 256 (XATTR_NAME_MAX)
356 * value 65536 (XATTR_SIZE_MAX)
358 #define MDS_EA_MAXREQSIZE 66288
361 * These are the maximum request and reply sizes (rounded up to 1 KB
362 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
364 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
365 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
366 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
369 * The update request includes all of updates from the create, which might
370 * include linkea (4K maxim), together with other updates, we set it to 9K:
371 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
373 #define MDS_OUT_MAXREQSIZE (9 * 1024)
374 #define MDS_OUT_MAXREPSIZE MDS_MAXREPSIZE
376 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
377 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
381 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
382 * However, we need to allocate a much larger buffer for it because LNet
383 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
384 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
385 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
386 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
387 * utilization is very low.
389 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
390 * reused until all requests fit in it have been processed and released,
391 * which means one long blocked request can prevent the rqbd be reused.
392 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
393 * from LNet with unused 65 KB, buffer utilization will be about 59%.
394 * Please check LU-2432 for details.
396 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
400 * MDS_OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
401 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
402 * extra bytes to each request buffer to improve buffer utilization rate.
404 #define MDS_OUT_BUFSIZE max(MDS_OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
407 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
408 #define FLD_MAXREQSIZE (160)
410 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
411 #define FLD_MAXREPSIZE (152)
412 #define FLD_BUFSIZE (1 << 12)
415 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
417 #define SEQ_MAXREQSIZE (160)
419 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
420 #define SEQ_MAXREPSIZE (152)
421 #define SEQ_BUFSIZE (1 << 12)
423 /** MGS threads must be >= 3, see bug 22458 comment #28 */
424 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
425 #define MGS_NTHRS_MAX 32
428 #define MGS_BUFSIZE (8 * 1024)
429 #define MGS_MAXREQSIZE (7 * 1024)
430 #define MGS_MAXREPSIZE (9 * 1024)
433 * OSS threads constants:
435 * Given 8 as factor and 64 as base threads number
438 * On 8-core server configured to 2 partitions, we will have
439 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
442 * On 32-core machine configured to 4 partitions, we will have
443 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
444 * will be 112 * 4 = 448.
447 * On 64-core machine configured to 4 partitions, we will have
448 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
449 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
450 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
451 * for each partition.
453 * So we can see that with these constants, threads number wil be at the
454 * similar level of old versions, unless the server has many cores.
456 /* depress threads factor for VM with small memory size */
457 #define OSS_THR_FACTOR min_t(int, 8, \
458 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
459 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
460 #define OSS_NTHRS_BASE 64
461 #define OSS_NTHRS_MAX 512
463 /* threads for handling "create" request */
464 #define OSS_CR_THR_FACTOR 1
465 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
466 #define OSS_CR_NTHRS_BASE 8
467 #define OSS_CR_NTHRS_MAX 64
470 * OST_IO_MAXREQSIZE ~=
471 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
472 * DT_MAX_BRW_PAGES * niobuf_remote
474 * - single object with 16 pages is 512 bytes
475 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
476 * - Must be a multiple of 1024
477 * - actual size is about 18K
479 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
480 sizeof(struct ptlrpc_body) + \
481 sizeof(struct obdo) + \
482 sizeof(struct obd_ioobj) + \
483 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
485 * FIEMAP request can be 4K+ for now
487 #define OST_MAXREQSIZE (5 * 1024)
488 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
489 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
491 #define OST_MAXREPSIZE (9 * 1024)
492 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
495 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
496 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
498 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
499 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
501 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
503 /* Macro to hide a typecast. */
504 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
507 * Structure to single define portal connection.
509 struct ptlrpc_connection {
510 /** linkage for connections hash table */
511 struct hlist_node c_hash;
512 /** Our own lnet nid for this connection */
514 /** Remote side nid for this connection */
515 lnet_process_id_t c_peer;
516 /** UUID of the other side */
517 struct obd_uuid c_remote_uuid;
518 /** reference counter for this connection */
522 /** Client definition for PortalRPC */
523 struct ptlrpc_client {
524 /** What lnet portal does this client send messages to by default */
525 __u32 cli_request_portal;
526 /** What portal do we expect replies on */
527 __u32 cli_reply_portal;
528 /** Name of the client */
532 /** state flags of requests */
533 /* XXX only ones left are those used by the bulk descs as well! */
534 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
535 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
537 #define REQ_MAX_ACK_LOCKS 8
539 union ptlrpc_async_args {
541 * Scratchpad for passing args to completion interpreter. Users
542 * cast to the struct of their choosing, and CLASSERT that this is
543 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
544 * a pointer to it here. The pointer_arg ensures this struct is at
545 * least big enough for that.
547 void *pointer_arg[11];
551 struct ptlrpc_request_set;
552 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
553 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
556 * Definition of request set structure.
557 * Request set is a list of requests (not necessary to the same target) that
558 * once populated with RPCs could be sent in parallel.
559 * There are two kinds of request sets. General purpose and with dedicated
560 * serving thread. Example of the latter is ptlrpcd set.
561 * For general purpose sets once request set started sending it is impossible
562 * to add new requests to such set.
563 * Provides a way to call "completion callbacks" when all requests in the set
566 struct ptlrpc_request_set {
567 atomic_t set_refcount;
568 /** number of in queue requests */
569 atomic_t set_new_count;
570 /** number of uncompleted requests */
571 atomic_t set_remaining;
572 /** wait queue to wait on for request events */
573 wait_queue_head_t set_waitq;
574 wait_queue_head_t *set_wakeup_ptr;
575 /** List of requests in the set */
576 struct list_head set_requests;
578 * List of completion callbacks to be called when the set is completed
579 * This is only used if \a set_interpret is NULL.
580 * Links struct ptlrpc_set_cbdata.
582 struct list_head set_cblist;
583 /** Completion callback, if only one. */
584 set_interpreter_func set_interpret;
585 /** opaq argument passed to completion \a set_interpret callback. */
588 * Lock for \a set_new_requests manipulations
589 * locked so that any old caller can communicate requests to
590 * the set holder who can then fold them into the lock-free set
592 spinlock_t set_new_req_lock;
593 /** List of new yet unsent requests. Only used with ptlrpcd now. */
594 struct list_head set_new_requests;
596 /** rq_status of requests that have been freed already */
598 /** Additional fields used by the flow control extension */
599 /** Maximum number of RPCs in flight */
600 int set_max_inflight;
601 /** Callback function used to generate RPCs */
602 set_producer_func set_producer;
603 /** opaq argument passed to the producer callback */
604 void *set_producer_arg;
608 * Description of a single ptrlrpc_set callback
610 struct ptlrpc_set_cbdata {
611 /** List linkage item */
612 struct list_head psc_item;
613 /** Pointer to interpreting function */
614 set_interpreter_func psc_interpret;
615 /** Opaq argument to pass to the callback */
619 struct ptlrpc_bulk_desc;
620 struct ptlrpc_service_part;
621 struct ptlrpc_service;
624 * ptlrpc callback & work item stuff
626 struct ptlrpc_cb_id {
627 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
628 void *cbid_arg; /* additional arg */
631 /** Maximum number of locks to fit into reply state */
632 #define RS_MAX_LOCKS 8
636 * Structure to define reply state on the server
637 * Reply state holds various reply message information. Also for "difficult"
638 * replies (rep-ack case) we store the state after sending reply and wait
639 * for the client to acknowledge the reception. In these cases locks could be
640 * added to the state for replay/failover consistency guarantees.
642 struct ptlrpc_reply_state {
643 /** Callback description */
644 struct ptlrpc_cb_id rs_cb_id;
645 /** Linkage for list of all reply states in a system */
646 struct list_head rs_list;
647 /** Linkage for list of all reply states on same export */
648 struct list_head rs_exp_list;
649 /** Linkage for list of all reply states for same obd */
650 struct list_head rs_obd_list;
652 struct list_head rs_debug_list;
654 /** A spinlock to protect the reply state flags */
656 /** Reply state flags */
657 unsigned long rs_difficult:1; /* ACK/commit stuff */
658 unsigned long rs_no_ack:1; /* no ACK, even for
659 difficult requests */
660 unsigned long rs_scheduled:1; /* being handled? */
661 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
662 unsigned long rs_handled:1; /* been handled yet? */
663 unsigned long rs_on_net:1; /* reply_out_callback pending? */
664 unsigned long rs_prealloc:1; /* rs from prealloc list */
665 unsigned long rs_committed:1;/* the transaction was committed
666 and the rs was dispatched
667 by ptlrpc_commit_replies */
668 /** Size of the state */
672 /** Transaction number */
676 struct obd_export *rs_export;
677 struct ptlrpc_service_part *rs_svcpt;
678 /** Lnet metadata handle for the reply */
679 lnet_handle_md_t rs_md_h;
680 atomic_t rs_refcount;
682 /** Context for the sevice thread */
683 struct ptlrpc_svc_ctx *rs_svc_ctx;
684 /** Reply buffer (actually sent to the client), encoded if needed */
685 struct lustre_msg *rs_repbuf; /* wrapper */
686 /** Size of the reply buffer */
687 int rs_repbuf_len; /* wrapper buf length */
688 /** Size of the reply message */
689 int rs_repdata_len; /* wrapper msg length */
691 * Actual reply message. Its content is encrupted (if needed) to
692 * produce reply buffer for actual sending. In simple case
693 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
695 struct lustre_msg *rs_msg; /* reply message */
697 /** Number of locks awaiting client ACK */
699 /** Handles of locks awaiting client reply ACK */
700 struct lustre_handle rs_locks[RS_MAX_LOCKS];
701 /** Lock modes of locks in \a rs_locks */
702 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
705 struct ptlrpc_thread;
709 RQ_PHASE_NEW = 0xebc0de00,
710 RQ_PHASE_RPC = 0xebc0de01,
711 RQ_PHASE_BULK = 0xebc0de02,
712 RQ_PHASE_INTERPRET = 0xebc0de03,
713 RQ_PHASE_COMPLETE = 0xebc0de04,
714 RQ_PHASE_UNREGISTERING = 0xebc0de05,
715 RQ_PHASE_UNDEFINED = 0xebc0de06
718 /** Type of request interpreter call-back */
719 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
720 struct ptlrpc_request *req,
724 * Definition of request pool structure.
725 * The pool is used to store empty preallocated requests for the case
726 * when we would actually need to send something without performing
727 * any allocations (to avoid e.g. OOM).
729 struct ptlrpc_request_pool {
730 /** Locks the list */
732 /** list of ptlrpc_request structs */
733 struct list_head prp_req_list;
734 /** Maximum message size that would fit into a rquest from this pool */
736 /** Function to allocate more requests for this pool */
737 void (*prp_populate)(struct ptlrpc_request_pool *, int);
746 * \defgroup nrs Network Request Scheduler
749 struct ptlrpc_nrs_policy;
750 struct ptlrpc_nrs_resource;
751 struct ptlrpc_nrs_request;
754 * NRS control operations.
756 * These are common for all policies.
758 enum ptlrpc_nrs_ctl {
760 * Not a valid opcode.
762 PTLRPC_NRS_CTL_INVALID,
764 * Activate the policy.
766 PTLRPC_NRS_CTL_START,
768 * Reserved for multiple primary policies, which may be a possibility
773 * Policies can start using opcodes from this value and onwards for
774 * their own purposes; the assigned value itself is arbitrary.
776 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
780 * ORR policy operations
783 NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
784 NRS_CTL_ORR_WR_QUANTUM,
785 NRS_CTL_ORR_RD_OFF_TYPE,
786 NRS_CTL_ORR_WR_OFF_TYPE,
787 NRS_CTL_ORR_RD_SUPP_REQ,
788 NRS_CTL_ORR_WR_SUPP_REQ,
792 * NRS policy operations.
794 * These determine the behaviour of a policy, and are called in response to
797 struct ptlrpc_nrs_pol_ops {
799 * Called during policy registration; this operation is optional.
801 * \param[in,out] policy The policy being initialized
803 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
805 * Called during policy unregistration; this operation is optional.
807 * \param[in,out] policy The policy being unregistered/finalized
809 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
811 * Called when activating a policy via lprocfs; policies allocate and
812 * initialize their resources here; this operation is optional.
814 * \param[in,out] policy The policy being started
816 * \see nrs_policy_start_locked()
818 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
820 * Called when deactivating a policy via lprocfs; policies deallocate
821 * their resources here; this operation is optional
823 * \param[in,out] policy The policy being stopped
825 * \see nrs_policy_stop0()
827 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
829 * Used for policy-specific operations; i.e. not generic ones like
830 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
831 * to an ioctl; this operation is optional.
833 * \param[in,out] policy The policy carrying out operation \a opc
834 * \param[in] opc The command operation being carried out
835 * \param[in,out] arg An generic buffer for communication between the
836 * user and the control operation
841 * \see ptlrpc_nrs_policy_control()
843 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
844 enum ptlrpc_nrs_ctl opc, void *arg);
847 * Called when obtaining references to the resources of the resource
848 * hierarchy for a request that has arrived for handling at the PTLRPC
849 * service. Policies should return -ve for requests they do not wish
850 * to handle. This operation is mandatory.
852 * \param[in,out] policy The policy we're getting resources for.
853 * \param[in,out] nrq The request we are getting resources for.
854 * \param[in] parent The parent resource of the resource being
855 * requested; set to NULL if none.
856 * \param[out] resp The resource is to be returned here; the
857 * fallback policy in an NRS head should
858 * \e always return a non-NULL pointer value.
859 * \param[in] moving_req When set, signifies that this is an attempt
860 * to obtain resources for a request being moved
861 * to the high-priority NRS head by
862 * ldlm_lock_reorder_req().
863 * This implies two things:
864 * 1. We are under obd_export::exp_rpc_lock and
865 * so should not sleep.
866 * 2. We should not perform non-idempotent or can
867 * skip performing idempotent operations that
868 * were carried out when resources were first
869 * taken for the request when it was initialized
870 * in ptlrpc_nrs_req_initialize().
872 * \retval 0, +ve The level of the returned resource in the resource
873 * hierarchy; currently only 0 (for a non-leaf resource)
874 * and 1 (for a leaf resource) are supported by the
878 * \see ptlrpc_nrs_req_initialize()
879 * \see ptlrpc_nrs_hpreq_add_nolock()
880 * \see ptlrpc_nrs_req_hp_move()
882 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
883 struct ptlrpc_nrs_request *nrq,
884 const struct ptlrpc_nrs_resource *parent,
885 struct ptlrpc_nrs_resource **resp,
888 * Called when releasing references taken for resources in the resource
889 * hierarchy for the request; this operation is optional.
891 * \param[in,out] policy The policy the resource belongs to
892 * \param[in] res The resource to be freed
894 * \see ptlrpc_nrs_req_finalize()
895 * \see ptlrpc_nrs_hpreq_add_nolock()
896 * \see ptlrpc_nrs_req_hp_move()
898 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
899 const struct ptlrpc_nrs_resource *res);
902 * Obtains a request for handling from the policy, and optionally
903 * removes the request from the policy; this operation is mandatory.
905 * \param[in,out] policy The policy to poll
906 * \param[in] peek When set, signifies that we just want to
907 * examine the request, and not handle it, so the
908 * request is not removed from the policy.
909 * \param[in] force When set, it will force a policy to return a
910 * request if it has one queued.
912 * \retval NULL No request available for handling
913 * \retval valid-pointer The request polled for handling
915 * \see ptlrpc_nrs_req_get_nolock()
917 struct ptlrpc_nrs_request *
918 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
921 * Called when attempting to add a request to a policy for later
922 * handling; this operation is mandatory.
924 * \param[in,out] policy The policy on which to enqueue \a nrq
925 * \param[in,out] nrq The request to enqueue
930 * \see ptlrpc_nrs_req_add_nolock()
932 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
933 struct ptlrpc_nrs_request *nrq);
935 * Removes a request from the policy's set of pending requests. Normally
936 * called after a request has been polled successfully from the policy
937 * for handling; this operation is mandatory.
939 * \param[in,out] policy The policy the request \a nrq belongs to
940 * \param[in,out] nrq The request to dequeue
942 * \see ptlrpc_nrs_req_del_nolock()
944 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
945 struct ptlrpc_nrs_request *nrq);
947 * Called after the request being carried out. Could be used for
948 * job/resource control; this operation is optional.
950 * \param[in,out] policy The policy which is stopping to handle request
952 * \param[in,out] nrq The request
954 * \pre spin_is_locked(&svcpt->scp_req_lock)
956 * \see ptlrpc_nrs_req_stop_nolock()
958 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
959 struct ptlrpc_nrs_request *nrq);
961 * Registers the policy's lprocfs interface with a PTLRPC service.
963 * \param[in] svc The service
968 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
970 * Unegisters the policy's lprocfs interface with a PTLRPC service.
972 * In cases of failed policy registration in
973 * \e ptlrpc_nrs_policy_register(), this function may be called for a
974 * service which has not registered the policy successfully, so
975 * implementations of this method should make sure their operations are
976 * safe in such cases.
978 * \param[in] svc The service
980 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
986 enum nrs_policy_flags {
988 * Fallback policy, use this flag only on a single supported policy per
989 * service. The flag cannot be used on policies that use
990 * \e PTLRPC_NRS_FL_REG_EXTERN
992 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
994 * Start policy immediately after registering.
996 PTLRPC_NRS_FL_REG_START = (1 << 1),
998 * This is a policy registering from a module different to the one NRS
999 * core ships in (currently ptlrpc).
1001 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1007 * Denotes whether an NRS instance is for handling normal or high-priority
1008 * RPCs, or whether an operation pertains to one or both of the NRS instances
1011 enum ptlrpc_nrs_queue_type {
1012 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1013 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1014 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1020 * A PTLRPC service has at least one NRS head instance for handling normal
1021 * priority RPCs, and may optionally have a second NRS head instance for
1022 * handling high-priority RPCs. Each NRS head maintains a list of available
1023 * policies, of which one and only one policy is acting as the fallback policy,
1024 * and optionally a different policy may be acting as the primary policy. For
1025 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1026 * enqueue the RPC using the primary policy (if any). The fallback policy is
1027 * used in the following cases:
1028 * - when there was no primary policy in the
1029 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1031 * - when the primary policy that was at the
1032 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1033 * RPC was initialized, denoted it did not wish, or for some other reason was
1034 * not able to handle the request, by returning a non-valid NRS resource
1036 * - when the primary policy that was at the
1037 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1038 * RPC was initialized, fails later during the request enqueueing stage.
1040 * \see nrs_resource_get_safe()
1041 * \see nrs_request_enqueue()
1044 spinlock_t nrs_lock;
1045 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1047 * List of registered policies
1049 struct list_head nrs_policy_list;
1051 * List of policies with queued requests. Policies that have any
1052 * outstanding requests are queued here, and this list is queried
1053 * in a round-robin manner from NRS core when obtaining a request
1054 * for handling. This ensures that requests from policies that at some
1055 * point transition away from the
1056 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1058 struct list_head nrs_policy_queued;
1060 * Service partition for this NRS head
1062 struct ptlrpc_service_part *nrs_svcpt;
1064 * Primary policy, which is the preferred policy for handling RPCs
1066 struct ptlrpc_nrs_policy *nrs_policy_primary;
1068 * Fallback policy, which is the backup policy for handling RPCs
1070 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1072 * This NRS head handles either HP or regular requests
1074 enum ptlrpc_nrs_queue_type nrs_queue_type;
1076 * # queued requests from all policies in this NRS head
1078 unsigned long nrs_req_queued;
1080 * # scheduled requests from all policies in this NRS head
1082 unsigned long nrs_req_started;
1084 * # policies on this NRS
1086 unsigned nrs_num_pols;
1088 * This NRS head is in progress of starting a policy
1090 unsigned nrs_policy_starting:1;
1092 * In progress of shutting down the whole NRS head; used during
1095 unsigned nrs_stopping:1;
1098 #define NRS_POL_NAME_MAX 16
1100 struct ptlrpc_nrs_pol_desc;
1103 * Service compatibility predicate; this determines whether a policy is adequate
1104 * for handling RPCs of a particular PTLRPC service.
1106 * XXX:This should give the same result during policy registration and
1107 * unregistration, and for all partitions of a service; so the result should not
1108 * depend on temporal service or other properties, that may influence the
1111 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1112 const struct ptlrpc_nrs_pol_desc *desc);
1114 struct ptlrpc_nrs_pol_conf {
1116 * Human-readable policy name
1118 char nc_name[NRS_POL_NAME_MAX];
1120 * NRS operations for this policy
1122 const struct ptlrpc_nrs_pol_ops *nc_ops;
1124 * Service compatibility predicate
1126 nrs_pol_desc_compat_t nc_compat;
1128 * Set for policies that support a single ptlrpc service, i.e. ones that
1129 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1130 * depicts the name of the single service that such policies are
1133 const char *nc_compat_svc_name;
1135 * Owner module for this policy descriptor; policies registering from a
1136 * different module to the one the NRS framework is held within
1137 * (currently ptlrpc), should set this field to THIS_MODULE.
1141 * Policy registration flags; a bitmast of \e nrs_policy_flags
1147 * NRS policy registering descriptor
1149 * Is used to hold a description of a policy that can be passed to NRS core in
1150 * order to register the policy with NRS heads in different PTLRPC services.
1152 struct ptlrpc_nrs_pol_desc {
1154 * Human-readable policy name
1156 char pd_name[NRS_POL_NAME_MAX];
1158 * Link into nrs_core::nrs_policies
1160 struct list_head pd_list;
1162 * NRS operations for this policy
1164 const struct ptlrpc_nrs_pol_ops *pd_ops;
1166 * Service compatibility predicate
1168 nrs_pol_desc_compat_t pd_compat;
1170 * Set for policies that are compatible with only one PTLRPC service.
1172 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1174 const char *pd_compat_svc_name;
1176 * Owner module for this policy descriptor.
1178 * We need to hold a reference to the module whenever we might make use
1179 * of any of the module's contents, i.e.
1180 * - If one or more instances of the policy are at a state where they
1181 * might be handling a request, i.e.
1182 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1183 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1184 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1185 * is taken on the module when
1186 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1187 * becomes 0, so that we hold only one reference to the module maximum
1190 * We do not need to hold a reference to the module, even though we
1191 * might use code and data from the module, in the following cases:
1192 * - During external policy registration, because this should happen in
1193 * the module's init() function, in which case the module is safe from
1194 * removal because a reference is being held on the module by the
1195 * kernel, and iirc kmod (and I guess module-init-tools also) will
1196 * serialize any racing processes properly anyway.
1197 * - During external policy unregistration, because this should happen
1198 * in a module's exit() function, and any attempts to start a policy
1199 * instance would need to take a reference on the module, and this is
1200 * not possible once we have reached the point where the exit()
1201 * handler is called.
1202 * - During service registration and unregistration, as service setup
1203 * and cleanup, and policy registration, unregistration and policy
1204 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1205 * as long as users adhere to the convention of registering policies
1206 * in init() and unregistering them in module exit() functions, there
1207 * should not be a race between these operations.
1208 * - During any policy-specific lprocfs operations, because a reference
1209 * is held by the kernel on a proc entry that has been entered by a
1210 * syscall, so as long as proc entries are removed during unregistration time,
1211 * then unregistration and lprocfs operations will be properly
1216 * Bitmask of \e nrs_policy_flags
1220 * # of references on this descriptor
1228 * Policies transition from one state to the other during their lifetime
1230 enum ptlrpc_nrs_pol_state {
1232 * Not a valid policy state.
1234 NRS_POL_STATE_INVALID,
1236 * Policies are at this state either at the start of their life, or
1237 * transition here when the user selects a different policy to act
1238 * as the primary one.
1240 NRS_POL_STATE_STOPPED,
1242 * Policy is progress of stopping
1244 NRS_POL_STATE_STOPPING,
1246 * Policy is in progress of starting
1248 NRS_POL_STATE_STARTING,
1250 * A policy is in this state in two cases:
1251 * - it is the fallback policy, which is always in this state.
1252 * - it has been activated by the user; i.e. it is the primary policy,
1254 NRS_POL_STATE_STARTED,
1258 * NRS policy information
1260 * Used for obtaining information for the status of a policy via lprocfs
1262 struct ptlrpc_nrs_pol_info {
1266 char pi_name[NRS_POL_NAME_MAX];
1268 * Current policy state
1270 enum ptlrpc_nrs_pol_state pi_state;
1272 * # RPCs enqueued for later dispatching by the policy
1276 * # RPCs started for dispatch by the policy
1278 long pi_req_started;
1280 * Is this a fallback policy?
1282 unsigned pi_fallback:1;
1288 * There is one instance of this for each policy in each NRS head of each
1289 * PTLRPC service partition.
1291 struct ptlrpc_nrs_policy {
1293 * Linkage into the NRS head's list of policies,
1294 * ptlrpc_nrs:nrs_policy_list
1296 struct list_head pol_list;
1298 * Linkage into the NRS head's list of policies with enqueued
1299 * requests ptlrpc_nrs:nrs_policy_queued
1301 struct list_head pol_list_queued;
1303 * Current state of this policy
1305 enum ptlrpc_nrs_pol_state pol_state;
1307 * Bitmask of nrs_policy_flags
1311 * # RPCs enqueued for later dispatching by the policy
1313 long pol_req_queued;
1315 * # RPCs started for dispatch by the policy
1317 long pol_req_started;
1319 * Usage Reference count taken on the policy instance
1323 * The NRS head this policy has been created at
1325 struct ptlrpc_nrs *pol_nrs;
1327 * Private policy data; varies by policy type
1331 * Policy descriptor for this policy instance.
1333 struct ptlrpc_nrs_pol_desc *pol_desc;
1339 * Resources are embedded into two types of NRS entities:
1340 * - Inside NRS policies, in the policy's private data in
1341 * ptlrpc_nrs_policy::pol_private
1342 * - In objects that act as prime-level scheduling entities in different NRS
1343 * policies; e.g. on a policy that performs round robin or similar order
1344 * scheduling across client NIDs, there would be one NRS resource per unique
1345 * client NID. On a policy which performs round robin scheduling across
1346 * backend filesystem objects, there would be one resource associated with
1347 * each of the backend filesystem objects partaking in the scheduling
1348 * performed by the policy.
1350 * NRS resources share a parent-child relationship, in which resources embedded
1351 * in policy instances are the parent entities, with all scheduling entities
1352 * a policy schedules across being the children, thus forming a simple resource
1353 * hierarchy. This hierarchy may be extended with one or more levels in the
1354 * future if the ability to have more than one primary policy is added.
1356 * Upon request initialization, references to the then active NRS policies are
1357 * taken and used to later handle the dispatching of the request with one of
1360 * \see nrs_resource_get_safe()
1361 * \see ptlrpc_nrs_req_add()
1363 struct ptlrpc_nrs_resource {
1365 * This NRS resource's parent; is NULL for resources embedded in NRS
1366 * policy instances; i.e. those are top-level ones.
1368 struct ptlrpc_nrs_resource *res_parent;
1370 * The policy associated with this resource.
1372 struct ptlrpc_nrs_policy *res_policy;
1385 * This policy is a logical wrapper around previous, non-NRS functionality.
1386 * It dispatches RPCs in the same order as they arrive from the network. This
1387 * policy is currently used as the fallback policy, and the only enabled policy
1388 * on all NRS heads of all PTLRPC service partitions.
1393 * Private data structure for the FIFO policy
1395 struct nrs_fifo_head {
1397 * Resource object for policy instance.
1399 struct ptlrpc_nrs_resource fh_res;
1401 * List of queued requests.
1403 struct list_head fh_list;
1405 * For debugging purposes.
1410 struct nrs_fifo_req {
1411 struct list_head fr_list;
1420 * CRR-N, Client Round Robin over NIDs
1425 * private data structure for CRR-N NRS
1427 struct nrs_crrn_net {
1428 struct ptlrpc_nrs_resource cn_res;
1429 cfs_binheap_t *cn_binheap;
1430 cfs_hash_t *cn_cli_hash;
1432 * Used when a new scheduling round commences, in order to synchronize
1433 * all clients with the new round number.
1437 * Determines the relevant ordering amongst request batches within a
1442 * Round Robin quantum; the maximum number of RPCs that each request
1443 * batch for each client can have in a scheduling round.
1449 * Object representing a client in CRR-N, as identified by its NID
1451 struct nrs_crrn_client {
1452 struct ptlrpc_nrs_resource cc_res;
1453 struct hlist_node cc_hnode;
1456 * The round number against which this client is currently scheduling
1461 * The sequence number used for requests scheduled by this client during
1462 * the current round number.
1467 * Round Robin quantum; the maximum number of RPCs the client is allowed
1468 * to schedule in a single batch of each round.
1472 * # of pending requests for this client, on all existing rounds
1478 * CRR-N NRS request definition
1480 struct nrs_crrn_req {
1482 * Round number for this request; shared with all other requests in the
1487 * Sequence number for this request; shared with all other requests in
1494 * CRR-N policy operations.
1498 * Read the RR quantum size of a CRR-N policy.
1500 NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
1502 * Write the RR quantum size of a CRR-N policy.
1504 NRS_CTL_CRRN_WR_QUANTUM,
1512 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1517 * Lower and upper byte offsets of a brw RPC
1519 struct nrs_orr_req_range {
1525 * RPC types supported by the ORR/TRR policies
1528 NOS_OST_READ = (1 << 0),
1529 NOS_OST_WRITE = (1 << 1),
1530 NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE),
1532 * Default value for policies.
1534 NOS_DFLT = NOS_OST_READ
1538 * As unique keys for grouping RPCs together, we use the object's OST FID for
1539 * the ORR policy, and the OST index for the TRR policy.
1541 * XXX: We waste some space for TRR policy instances by using a union, but it
1542 * allows to consolidate some of the code between ORR and TRR, and these
1543 * policies will probably eventually merge into one anyway.
1545 struct nrs_orr_key {
1547 /** object FID for ORR */
1548 struct lu_fid ok_fid;
1549 /** OST index for TRR */
1555 * The largest base string for unique hash/slab object names is
1556 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1557 * id number, so this _should_ be more than enough for the maximum number of
1558 * CPTs on any system. If it does happen that this statement is incorrect,
1559 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1560 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1561 * will hopefully not go unnoticed.
1563 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1566 * private data structure for ORR and TRR NRS
1568 struct nrs_orr_data {
1569 struct ptlrpc_nrs_resource od_res;
1570 cfs_binheap_t *od_binheap;
1571 cfs_hash_t *od_obj_hash;
1572 struct kmem_cache *od_cache;
1574 * Used when a new scheduling round commences, in order to synchronize
1575 * all object or OST batches with the new round number.
1579 * Determines the relevant ordering amongst request batches within a
1584 * RPC types that are currently supported.
1586 enum nrs_orr_supp od_supp;
1588 * Round Robin quantum; the maxium number of RPCs that each request
1589 * batch for each object or OST can have in a scheduling round.
1593 * Whether to use physical disk offsets or logical file offsets.
1597 * XXX: We need to provide a persistently allocated string to hold
1598 * unique object names for this policy, since in currently supported
1599 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1600 * to the name string provided. kstrdup() is used in the version of
1601 * kmeme_cache_create() in current Linux mainline, so we may be able to
1602 * remove this in the future.
1604 char od_objname[NRS_ORR_OBJ_NAME_MAX];
1608 * Represents a backend-fs object or OST in the ORR and TRR policies
1611 struct nrs_orr_object {
1612 struct ptlrpc_nrs_resource oo_res;
1613 struct hlist_node oo_hnode;
1615 * The round number against which requests are being scheduled for this
1620 * The sequence number used for requests scheduled for this object or
1621 * OST during the current round number.
1625 * The key of the object or OST for which this structure instance is
1628 struct nrs_orr_key oo_key;
1631 * Round Robin quantum; the maximum number of RPCs that are allowed to
1632 * be scheduled for the object or OST in a single batch of each round.
1636 * # of pending requests for this object or OST, on all existing rounds
1642 * ORR/TRR NRS request definition
1644 struct nrs_orr_req {
1646 * The offset range this request covers
1648 struct nrs_orr_req_range or_range;
1650 * Round number for this request; shared with all other requests in the
1655 * Sequence number for this request; shared with all other requests in
1660 * For debugging purposes.
1662 struct nrs_orr_key or_key;
1664 * An ORR policy instance has filled in request information while
1665 * enqueueing the request on the service partition's regular NRS head.
1667 unsigned int or_orr_set:1;
1669 * A TRR policy instance has filled in request information while
1670 * enqueueing the request on the service partition's regular NRS head.
1672 unsigned int or_trr_set:1;
1674 * Request offset ranges have been filled in with logical offset
1677 unsigned int or_logical_set:1;
1679 * Request offset ranges have been filled in with physical offset
1682 unsigned int or_physical_set:1;
1690 * Instances of this object exist embedded within ptlrpc_request; the main
1691 * purpose of this object is to hold references to the request's resources
1692 * for the lifetime of the request, and to hold properties that policies use
1693 * use for determining the request's scheduling priority.
1695 struct ptlrpc_nrs_request {
1697 * The request's resource hierarchy.
1699 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1701 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1702 * policy that was used to enqueue the request.
1704 * \see nrs_request_enqueue()
1706 unsigned nr_res_idx;
1707 unsigned nr_initialized:1;
1708 unsigned nr_enqueued:1;
1709 unsigned nr_started:1;
1710 unsigned nr_finalized:1;
1711 cfs_binheap_node_t nr_node;
1714 * Policy-specific fields, used for determining a request's scheduling
1715 * priority, and other supporting functionality.
1719 * Fields for the FIFO policy
1721 struct nrs_fifo_req fifo;
1723 * CRR-N request defintion
1725 struct nrs_crrn_req crr;
1726 /** ORR and TRR share the same request definition */
1727 struct nrs_orr_req orr;
1730 * Externally-registering policies may want to use this to allocate
1731 * their own request properties.
1739 * Basic request prioritization operations structure.
1740 * The whole idea is centered around locks and RPCs that might affect locks.
1741 * When a lock is contended we try to give priority to RPCs that might lead
1742 * to fastest release of that lock.
1743 * Currently only implemented for OSTs only in a way that makes all
1744 * IO and truncate RPCs that are coming from a locked region where a lock is
1745 * contended a priority over other requests.
1747 struct ptlrpc_hpreq_ops {
1749 * Check if the lock handle of the given lock is the same as
1750 * taken from the request.
1752 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1754 * Check if the request is a high priority one.
1756 int (*hpreq_check)(struct ptlrpc_request *);
1758 * Called after the request has been handled.
1760 void (*hpreq_fini)(struct ptlrpc_request *);
1764 * Represents remote procedure call.
1766 * This is a staple structure used by everybody wanting to send a request
1769 struct ptlrpc_request {
1770 /* Request type: one of PTL_RPC_MSG_* */
1772 /** Result of request processing */
1775 * Linkage item through which this request is included into
1776 * sending/delayed lists on client and into rqbd list on server
1778 struct list_head rq_list;
1780 * Server side list of incoming unserved requests sorted by arrival
1781 * time. Traversed from time to time to notice about to expire
1782 * requests and sent back "early replies" to clients to let them
1783 * know server is alive and well, just very busy to service their
1786 struct list_head rq_timed_list;
1787 /** server-side history, used for debuging purposes. */
1788 struct list_head rq_history_list;
1789 /** server-side per-export list */
1790 struct list_head rq_exp_list;
1791 /** server-side hp handlers */
1792 struct ptlrpc_hpreq_ops *rq_ops;
1794 /** initial thread servicing this request */
1795 struct ptlrpc_thread *rq_svc_thread;
1797 /** history sequence # */
1798 __u64 rq_history_seq;
1802 /** stub for NRS request */
1803 struct ptlrpc_nrs_request rq_nrq;
1805 /** the index of service's srv_at_array into which request is linked */
1807 /** Lock to protect request flags and some other important bits, like
1811 /** client-side flags are serialized by rq_lock */
1812 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1813 rq_timedout:1, rq_resend:1, rq_restart:1,
1815 * when ->rq_replay is set, request is kept by the client even
1816 * after server commits corresponding transaction. This is
1817 * used for operations that require sequence of multiple
1818 * requests to be replayed. The only example currently is file
1819 * open/close. When last request in such a sequence is
1820 * committed, ->rq_replay is cleared on all requests in the
1824 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1825 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1826 rq_early:1, rq_must_unlink:1,
1827 rq_memalloc:1, /* req originated from "kswapd" */
1828 /* server-side flags */
1829 rq_packed_final:1, /* packed final reply */
1830 rq_hp:1, /* high priority RPC */
1831 rq_at_linked:1, /* link into service's srv_at_array */
1832 rq_reply_truncate:1,
1834 /* whether the "rq_set" is a valid one */
1836 rq_generation_set:1,
1837 /* do not resend request on -EINPROGRESS */
1838 rq_no_retry_einprogress:1,
1839 /* allow the req to be sent if the import is in recovery
1843 unsigned int rq_nr_resend;
1845 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1846 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1847 atomic_t rq_refcount;/* client-side refcount for SENT race,
1848 server-side refcounf for multiple replies */
1850 /** Portal to which this request would be sent */
1851 short rq_request_portal; /* XXX FIXME bug 249 */
1852 /** Portal where to wait for reply and where reply would be sent */
1853 short rq_reply_portal; /* XXX FIXME bug 249 */
1857 * !rq_truncate : # reply bytes actually received,
1858 * rq_truncate : required repbuf_len for resend
1860 int rq_nob_received;
1861 /** Request length */
1865 /** Request message - what client sent */
1866 struct lustre_msg *rq_reqmsg;
1867 /** Reply message - server response */
1868 struct lustre_msg *rq_repmsg;
1869 /** Transaction number */
1874 * List item to for replay list. Not yet commited requests get linked
1876 * Also see \a rq_replay comment above.
1878 struct list_head rq_replay_list;
1881 * security and encryption data
1883 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1884 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1885 struct list_head rq_ctx_chain; /**< link to waited ctx */
1887 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1888 enum lustre_sec_part rq_sp_from;
1890 /* client/server security flags */
1892 rq_ctx_init:1, /* context initiation */
1893 rq_ctx_fini:1, /* context destroy */
1894 rq_bulk_read:1, /* request bulk read */
1895 rq_bulk_write:1, /* request bulk write */
1896 /* server authentication flags */
1897 rq_auth_gss:1, /* authenticated by gss */
1898 rq_auth_remote:1, /* authed as remote user */
1899 rq_auth_usr_root:1, /* authed as root */
1900 rq_auth_usr_mdt:1, /* authed as mdt */
1901 rq_auth_usr_ost:1, /* authed as ost */
1902 /* security tfm flags */
1905 /* doesn't expect reply FIXME */
1907 rq_pill_init:1; /* pill initialized */
1909 uid_t rq_auth_uid; /* authed uid */
1910 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1912 /* (server side), pointed directly into req buffer */
1913 struct ptlrpc_user_desc *rq_user_desc;
1915 /* various buffer pointers */
1916 struct lustre_msg *rq_reqbuf; /* req wrapper */
1917 char *rq_repbuf; /* rep buffer */
1918 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1919 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1920 int rq_reqbuf_len; /* req wrapper buf len */
1921 int rq_reqdata_len; /* req wrapper msg len */
1922 int rq_repbuf_len; /* rep buffer len */
1923 int rq_repdata_len; /* rep wrapper msg len */
1924 int rq_clrbuf_len; /* only in priv mode */
1925 int rq_clrdata_len; /* only in priv mode */
1927 /** early replies go to offset 0, regular replies go after that */
1928 unsigned int rq_reply_off;
1932 /** Fields that help to see if request and reply were swabbed or not */
1933 __u32 rq_req_swab_mask;
1934 __u32 rq_rep_swab_mask;
1936 /** What was import generation when this request was sent */
1937 int rq_import_generation;
1938 enum lustre_imp_state rq_send_state;
1940 /** how many early replies (for stats) */
1943 /** client+server request */
1944 lnet_handle_md_t rq_req_md_h;
1945 struct ptlrpc_cb_id rq_req_cbid;
1946 /** optional time limit for send attempts */
1947 cfs_duration_t rq_delay_limit;
1948 /** time request was first queued */
1949 cfs_time_t rq_queued_time;
1951 /* server-side... */
1952 /** request arrival time */
1953 struct timeval rq_arrival_time;
1954 /** separated reply state */
1955 struct ptlrpc_reply_state *rq_reply_state;
1956 /** incoming request buffer */
1957 struct ptlrpc_request_buffer_desc *rq_rqbd;
1959 /** client-only incoming reply */
1960 lnet_handle_md_t rq_reply_md_h;
1961 wait_queue_head_t rq_reply_waitq;
1962 struct ptlrpc_cb_id rq_reply_cbid;
1966 /** Peer description (the other side) */
1967 lnet_process_id_t rq_peer;
1968 /** Server-side, export on which request was received */
1969 struct obd_export *rq_export;
1970 /** Client side, import where request is being sent */
1971 struct obd_import *rq_import;
1973 /** Replay callback, called after request is replayed at recovery */
1974 void (*rq_replay_cb)(struct ptlrpc_request *);
1976 * Commit callback, called when request is committed and about to be
1979 void (*rq_commit_cb)(struct ptlrpc_request *);
1980 /** Opaq data for replay and commit callbacks. */
1983 /** For bulk requests on client only: bulk descriptor */
1984 struct ptlrpc_bulk_desc *rq_bulk;
1986 /** client outgoing req */
1988 * when request/reply sent (secs), or time when request should be sent
1991 /** time for request really sent out */
1992 time_t rq_real_sent;
1994 /** when request must finish. volatile
1995 * so that servers' early reply updates to the deadline aren't
1996 * kept in per-cpu cache */
1997 volatile time_t rq_deadline;
1998 /** when req reply unlink must finish. */
1999 time_t rq_reply_deadline;
2000 /** when req bulk unlink must finish. */
2001 time_t rq_bulk_deadline;
2003 * service time estimate (secs)
2004 * If the requestsis not served by this time, it is marked as timed out.
2008 /** Multi-rpc bits */
2009 /** Per-request waitq introduced by bug 21938 for recovery waiting */
2010 wait_queue_head_t rq_set_waitq;
2011 /** Link item for request set lists */
2012 struct list_head rq_set_chain;
2013 /** Link back to the request set */
2014 struct ptlrpc_request_set *rq_set;
2015 /** Async completion handler, called when reply is received */
2016 ptlrpc_interpterer_t rq_interpret_reply;
2017 /** Async completion context */
2018 union ptlrpc_async_args rq_async_args;
2020 /** Pool if request is from preallocated list */
2021 struct ptlrpc_request_pool *rq_pool;
2023 struct lu_context rq_session;
2024 struct lu_context rq_recov_session;
2026 /** request format description */
2027 struct req_capsule rq_pill;
2031 * Call completion handler for rpc if any, return it's status or original
2032 * rc if there was no handler defined for this request.
2034 static inline int ptlrpc_req_interpret(const struct lu_env *env,
2035 struct ptlrpc_request *req, int rc)
2037 if (req->rq_interpret_reply != NULL) {
2038 req->rq_status = req->rq_interpret_reply(env, req,
2039 &req->rq_async_args,
2041 return req->rq_status;
2049 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
2050 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
2051 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
2052 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
2053 struct ptlrpc_nrs_pol_info *info);
2056 * Can the request be moved from the regular NRS head to the high-priority NRS
2057 * head (of the same PTLRPC service partition), if any?
2059 * For a reliable result, this should be checked under svcpt->scp_req lock.
2061 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
2063 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
2066 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
2067 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
2068 * to make sure it has not been scheduled yet (analogous to previous
2069 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
2071 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
2076 * Returns 1 if request buffer at offset \a index was already swabbed
2078 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
2080 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2081 return req->rq_req_swab_mask & (1 << index);
2085 * Returns 1 if request reply buffer at offset \a index was already swabbed
2087 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
2089 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2090 return req->rq_rep_swab_mask & (1 << index);
2094 * Returns 1 if request needs to be swabbed into local cpu byteorder
2096 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
2098 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2102 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
2104 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
2106 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2110 * Mark request buffer at offset \a index that it was already swabbed
2112 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
2114 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2115 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
2116 req->rq_req_swab_mask |= 1 << index;
2120 * Mark request reply buffer at offset \a index that it was already swabbed
2122 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
2124 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2125 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
2126 req->rq_rep_swab_mask |= 1 << index;
2130 * Convert numerical request phase value \a phase into text string description
2132 static inline const char *
2133 ptlrpc_phase2str(enum rq_phase phase)
2142 case RQ_PHASE_INTERPRET:
2144 case RQ_PHASE_COMPLETE:
2146 case RQ_PHASE_UNREGISTERING:
2147 return "Unregistering";
2154 * Convert numerical request phase of the request \a req into text stringi
2157 static inline const char *
2158 ptlrpc_rqphase2str(struct ptlrpc_request *req)
2160 return ptlrpc_phase2str(req->rq_phase);
2164 * Debugging functions and helpers to print request structure into debug log
2167 /* Spare the preprocessor, spoil the bugs. */
2168 #define FLAG(field, str) (field ? str : "")
2170 /** Convert bit flags into a string */
2171 #define DEBUG_REQ_FLAGS(req) \
2172 ptlrpc_rqphase2str(req), \
2173 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
2174 FLAG(req->rq_err, "E"), \
2175 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
2176 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
2177 FLAG(req->rq_no_resend, "N"), \
2178 FLAG(req->rq_waiting, "W"), \
2179 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
2180 FLAG(req->rq_committed, "M")
2182 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
2184 void _debug_req(struct ptlrpc_request *req,
2185 struct libcfs_debug_msg_data *data, const char *fmt, ...)
2186 __attribute__ ((format (printf, 3, 4)));
2189 * Helper that decides if we need to print request accordig to current debug
2192 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
2194 CFS_CHECK_STACK(msgdata, mask, cdls); \
2196 if (((mask) & D_CANTMASK) != 0 || \
2197 ((libcfs_debug & (mask)) != 0 && \
2198 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
2199 _debug_req((req), msgdata, fmt, ##a); \
2203 * This is the debug print function you need to use to print request sturucture
2204 * content into lustre debug log.
2205 * for most callers (level is a constant) this is resolved at compile time */
2206 #define DEBUG_REQ(level, req, fmt, args...) \
2208 if ((level) & (D_ERROR | D_WARNING)) { \
2209 static cfs_debug_limit_state_t cdls; \
2210 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
2211 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
2213 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
2214 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
2220 * Structure that defines a single page of a bulk transfer
2222 struct ptlrpc_bulk_page {
2223 /** Linkage to list of pages in a bulk */
2224 struct list_head bp_link;
2226 * Number of bytes in a page to transfer starting from \a bp_pageoffset
2229 /** offset within a page */
2231 /** The page itself */
2232 struct page *bp_page;
2235 #define BULK_GET_SOURCE 0
2236 #define BULK_PUT_SINK 1
2237 #define BULK_GET_SINK 2
2238 #define BULK_PUT_SOURCE 3
2241 * Definition of bulk descriptor.
2242 * Bulks are special "Two phase" RPCs where initial request message
2243 * is sent first and it is followed bt a transfer (o receiving) of a large
2244 * amount of data to be settled into pages referenced from the bulk descriptors.
2245 * Bulks transfers (the actual data following the small requests) are done
2246 * on separate LNet portals.
2247 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2248 * Another user is readpage for MDT.
2250 struct ptlrpc_bulk_desc {
2251 /** completed with failure */
2252 unsigned long bd_failure:1;
2253 /** {put,get}{source,sink} */
2254 unsigned long bd_type:2;
2256 unsigned long bd_registered:1;
2257 /** For serialization with callback */
2259 /** Import generation when request for this bulk was sent */
2260 int bd_import_generation;
2261 /** LNet portal for this bulk */
2263 /** Server side - export this bulk created for */
2264 struct obd_export *bd_export;
2265 /** Client side - import this bulk was sent on */
2266 struct obd_import *bd_import;
2267 /** Back pointer to the request */
2268 struct ptlrpc_request *bd_req;
2269 wait_queue_head_t bd_waitq; /* server side only WQ */
2270 int bd_iov_count; /* # entries in bd_iov */
2271 int bd_max_iov; /* allocated size of bd_iov */
2272 int bd_nob; /* # bytes covered */
2273 int bd_nob_transferred; /* # bytes GOT/PUT */
2277 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2278 lnet_nid_t bd_sender; /* stash event::sender */
2279 int bd_md_count; /* # valid entries in bd_mds */
2280 int bd_md_max_brw; /* max entries in bd_mds */
2281 /** array of associated MDs */
2282 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2285 * encrypt iov, size is either 0 or bd_iov_count.
2287 lnet_kiov_t *bd_enc_iov;
2289 lnet_kiov_t bd_iov[0];
2293 SVC_STOPPED = 1 << 0,
2294 SVC_STOPPING = 1 << 1,
2295 SVC_STARTING = 1 << 2,
2296 SVC_RUNNING = 1 << 3,
2298 SVC_SIGNAL = 1 << 5,
2301 #define PTLRPC_THR_NAME_LEN 32
2303 * Definition of server service thread structure
2305 struct ptlrpc_thread {
2307 * List of active threads in svc->srv_threads
2309 struct list_head t_link;
2311 * thread-private data (preallocated memory)
2316 * service thread index, from ptlrpc_start_threads
2320 * service thread pid
2324 * put watchdog in the structure per thread b=14840
2326 * Lustre watchdog is removed for client in the hope
2327 * of a generic watchdog can be merged in kernel.
2328 * When that happens, we should add below back.
2330 * struct lc_watchdog *t_watchdog;
2333 * the svc this thread belonged to b=18582
2335 struct ptlrpc_service_part *t_svcpt;
2336 wait_queue_head_t t_ctl_waitq;
2337 struct lu_env *t_env;
2338 char t_name[PTLRPC_THR_NAME_LEN];
2341 static inline int thread_is_init(struct ptlrpc_thread *thread)
2343 return thread->t_flags == 0;
2346 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2348 return !!(thread->t_flags & SVC_STOPPED);
2351 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2353 return !!(thread->t_flags & SVC_STOPPING);
2356 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2358 return !!(thread->t_flags & SVC_STARTING);
2361 static inline int thread_is_running(struct ptlrpc_thread *thread)
2363 return !!(thread->t_flags & SVC_RUNNING);
2366 static inline int thread_is_event(struct ptlrpc_thread *thread)
2368 return !!(thread->t_flags & SVC_EVENT);
2371 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2373 return !!(thread->t_flags & SVC_SIGNAL);
2376 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2378 thread->t_flags &= ~flags;
2381 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2383 thread->t_flags = flags;
2386 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2388 thread->t_flags |= flags;
2391 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2394 if (thread->t_flags & flags) {
2395 thread->t_flags &= ~flags;
2402 * Request buffer descriptor structure.
2403 * This is a structure that contains one posted request buffer for service.
2404 * Once data land into a buffer, event callback creates actual request and
2405 * notifies wakes one of the service threads to process new incoming request.
2406 * More than one request can fit into the buffer.
2408 struct ptlrpc_request_buffer_desc {
2409 /** Link item for rqbds on a service */
2410 struct list_head rqbd_list;
2411 /** History of requests for this buffer */
2412 struct list_head rqbd_reqs;
2413 /** Back pointer to service for which this buffer is registered */
2414 struct ptlrpc_service_part *rqbd_svcpt;
2415 /** LNet descriptor */
2416 lnet_handle_md_t rqbd_md_h;
2418 /** The buffer itself */
2420 struct ptlrpc_cb_id rqbd_cbid;
2422 * This "embedded" request structure is only used for the
2423 * last request to fit into the buffer
2425 struct ptlrpc_request rqbd_req;
2428 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2430 struct ptlrpc_service_ops {
2432 * if non-NULL called during thread creation (ptlrpc_start_thread())
2433 * to initialize service specific per-thread state.
2435 int (*so_thr_init)(struct ptlrpc_thread *thr);
2437 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2438 * destruct state created by ->srv_init().
2440 void (*so_thr_done)(struct ptlrpc_thread *thr);
2442 * Handler function for incoming requests for this service
2444 int (*so_req_handler)(struct ptlrpc_request *req);
2446 * function to determine priority of the request, it's called
2447 * on every new request
2449 int (*so_hpreq_handler)(struct ptlrpc_request *);
2451 * service-specific print fn
2453 void (*so_req_printer)(void *, struct ptlrpc_request *);
2456 #ifndef __cfs_cacheline_aligned
2457 /* NB: put it here for reducing patche dependence */
2458 # define __cfs_cacheline_aligned
2462 * How many high priority requests to serve before serving one normal
2465 #define PTLRPC_SVC_HP_RATIO 10
2468 * Definition of PortalRPC service.
2469 * The service is listening on a particular portal (like tcp port)
2470 * and perform actions for a specific server like IO service for OST
2471 * or general metadata service for MDS.
2473 struct ptlrpc_service {
2474 /** serialize /proc operations */
2475 spinlock_t srv_lock;
2476 /** most often accessed fields */
2477 /** chain thru all services */
2478 struct list_head srv_list;
2479 /** service operations table */
2480 struct ptlrpc_service_ops srv_ops;
2481 /** only statically allocated strings here; we don't clean them */
2483 /** only statically allocated strings here; we don't clean them */
2484 char *srv_thread_name;
2485 /** service thread list */
2486 struct list_head srv_threads;
2487 /** threads # should be created for each partition on initializing */
2488 int srv_nthrs_cpt_init;
2489 /** limit of threads number for each partition */
2490 int srv_nthrs_cpt_limit;
2491 /** Root of /proc dir tree for this service */
2492 proc_dir_entry_t *srv_procroot;
2493 /** Pointer to statistic data for this service */
2494 struct lprocfs_stats *srv_stats;
2495 /** # hp per lp reqs to handle */
2496 int srv_hpreq_ratio;
2497 /** biggest request to receive */
2498 int srv_max_req_size;
2499 /** biggest reply to send */
2500 int srv_max_reply_size;
2501 /** size of individual buffers */
2503 /** # buffers to allocate in 1 group */
2504 int srv_nbuf_per_group;
2505 /** Local portal on which to receive requests */
2506 __u32 srv_req_portal;
2507 /** Portal on the client to send replies to */
2508 __u32 srv_rep_portal;
2510 * Tags for lu_context associated with this thread, see struct
2514 /** soft watchdog timeout multiplier */
2515 int srv_watchdog_factor;
2516 /** under unregister_service */
2517 unsigned srv_is_stopping:1;
2519 /** max # request buffers in history per partition */
2520 int srv_hist_nrqbds_cpt_max;
2521 /** number of CPTs this service bound on */
2523 /** CPTs array this service bound on */
2525 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2527 /** CPT table this service is running over */
2528 struct cfs_cpt_table *srv_cptable;
2530 * partition data for ptlrpc service
2532 struct ptlrpc_service_part *srv_parts[0];
2536 * Definition of PortalRPC service partition data.
2537 * Although a service only has one instance of it right now, but we
2538 * will have multiple instances very soon (instance per CPT).
2540 * it has four locks:
2542 * serialize operations on rqbd and requests waiting for preprocess
2544 * serialize operations active requests sent to this portal
2546 * serialize adaptive timeout stuff
2548 * serialize operations on RS list (reply states)
2550 * We don't have any use-case to take two or more locks at the same time
2551 * for now, so there is no lock order issue.
2553 struct ptlrpc_service_part {
2554 /** back reference to owner */
2555 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2556 /* CPT id, reserved */
2558 /** always increasing number */
2560 /** # of starting threads */
2561 int scp_nthrs_starting;
2562 /** # of stopping threads, reserved for shrinking threads */
2563 int scp_nthrs_stopping;
2564 /** # running threads */
2565 int scp_nthrs_running;
2566 /** service threads list */
2567 struct list_head scp_threads;
2570 * serialize the following fields, used for protecting
2571 * rqbd list and incoming requests waiting for preprocess,
2572 * threads starting & stopping are also protected by this lock.
2574 spinlock_t scp_lock __cfs_cacheline_aligned;
2575 /** total # req buffer descs allocated */
2576 int scp_nrqbds_total;
2577 /** # posted request buffers for receiving */
2578 int scp_nrqbds_posted;
2579 /** in progress of allocating rqbd */
2580 int scp_rqbd_allocating;
2581 /** # incoming reqs */
2582 int scp_nreqs_incoming;
2583 /** request buffers to be reposted */
2584 struct list_head scp_rqbd_idle;
2585 /** req buffers receiving */
2586 struct list_head scp_rqbd_posted;
2587 /** incoming reqs */
2588 struct list_head scp_req_incoming;
2589 /** timeout before re-posting reqs, in tick */
2590 cfs_duration_t scp_rqbd_timeout;
2592 * all threads sleep on this. This wait-queue is signalled when new
2593 * incoming request arrives and when difficult reply has to be handled.
2595 wait_queue_head_t scp_waitq;
2597 /** request history */
2598 struct list_head scp_hist_reqs;
2599 /** request buffer history */
2600 struct list_head scp_hist_rqbds;
2601 /** # request buffers in history */
2602 int scp_hist_nrqbds;
2603 /** sequence number for request */
2605 /** highest seq culled from history */
2606 __u64 scp_hist_seq_culled;
2609 * serialize the following fields, used for processing requests
2610 * sent to this portal
2612 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2613 /** # reqs in either of the NRS heads below */
2614 /** # reqs being served */
2615 int scp_nreqs_active;
2616 /** # HPreqs being served */
2617 int scp_nhreqs_active;
2618 /** # hp requests handled */
2621 /** NRS head for regular requests */
2622 struct ptlrpc_nrs scp_nrs_reg;
2623 /** NRS head for HP requests; this is only valid for services that can
2624 * handle HP requests */
2625 struct ptlrpc_nrs *scp_nrs_hp;
2630 * serialize the following fields, used for changes on
2633 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2634 /** estimated rpc service time */
2635 struct adaptive_timeout scp_at_estimate;
2636 /** reqs waiting for replies */
2637 struct ptlrpc_at_array scp_at_array;
2638 /** early reply timer */
2639 timer_list_t scp_at_timer;
2641 cfs_time_t scp_at_checktime;
2642 /** check early replies */
2643 unsigned scp_at_check;
2647 * serialize the following fields, used for processing
2648 * replies for this portal
2650 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2651 /** all the active replies */
2652 struct list_head scp_rep_active;
2653 /** List of free reply_states */
2654 struct list_head scp_rep_idle;
2655 /** waitq to run, when adding stuff to srv_free_rs_list */
2656 wait_queue_head_t scp_rep_waitq;
2657 /** # 'difficult' replies */
2658 atomic_t scp_nreps_difficult;
2661 #define ptlrpc_service_for_each_part(part, i, svc) \
2663 i < (svc)->srv_ncpts && \
2664 (svc)->srv_parts != NULL && \
2665 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2668 * Declaration of ptlrpcd control structure
2670 struct ptlrpcd_ctl {
2672 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2674 unsigned long pc_flags;
2676 * Thread lock protecting structure fields.
2682 struct completion pc_starting;
2686 struct completion pc_finishing;
2688 * Thread requests set.
2690 struct ptlrpc_request_set *pc_set;
2692 * Thread name used in cfs_daemonize()
2696 * Environment for request interpreters to run in.
2698 struct lu_env pc_env;
2700 * Index of ptlrpcd thread in the array.
2704 * Number of the ptlrpcd's partners.
2708 * Pointer to the array of partners' ptlrpcd_ctl structure.
2710 struct ptlrpcd_ctl **pc_partners;
2712 * Record the partner index to be processed next.
2717 /* Bits for pc_flags */
2718 enum ptlrpcd_ctl_flags {
2720 * Ptlrpc thread start flag.
2722 LIOD_START = 1 << 0,
2724 * Ptlrpc thread stop flag.
2728 * Ptlrpc thread force flag (only stop force so far).
2729 * This will cause aborting any inflight rpcs handled
2730 * by thread if LIOD_STOP is specified.
2732 LIOD_FORCE = 1 << 2,
2734 * This is a recovery ptlrpc thread.
2736 LIOD_RECOVERY = 1 << 3,
2738 * The ptlrpcd is bound to some CPU core.
2747 * Service compatibility function; the policy is compatible with all services.
2749 * \param[in] svc The service the policy is attempting to register with.
2750 * \param[in] desc The policy descriptor
2752 * \retval true The policy is compatible with the service
2754 * \see ptlrpc_nrs_pol_desc::pd_compat()
2756 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2757 const struct ptlrpc_nrs_pol_desc *desc)
2763 * Service compatibility function; the policy is compatible with only a specific
2764 * service which is identified by its human-readable name at
2765 * ptlrpc_service::srv_name.
2767 * \param[in] svc The service the policy is attempting to register with.
2768 * \param[in] desc The policy descriptor
2770 * \retval false The policy is not compatible with the service
2771 * \retval true The policy is compatible with the service
2773 * \see ptlrpc_nrs_pol_desc::pd_compat()
2775 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2776 const struct ptlrpc_nrs_pol_desc *desc)
2778 LASSERT(desc->pd_compat_svc_name != NULL);
2779 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2784 /* ptlrpc/events.c */
2785 extern lnet_handle_eq_t ptlrpc_eq_h;
2786 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2787 lnet_process_id_t *peer, lnet_nid_t *self);
2789 * These callbacks are invoked by LNet when something happened to
2793 extern void request_out_callback(lnet_event_t *ev);
2794 extern void reply_in_callback(lnet_event_t *ev);
2795 extern void client_bulk_callback(lnet_event_t *ev);
2796 extern void request_in_callback(lnet_event_t *ev);
2797 extern void reply_out_callback(lnet_event_t *ev);
2800 /* ptlrpc/connection.c */
2801 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2803 struct obd_uuid *uuid);
2804 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2805 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2806 int ptlrpc_connection_init(void);
2807 void ptlrpc_connection_fini(void);
2808 extern lnet_pid_t ptl_get_pid(void);
2810 /* ptlrpc/niobuf.c */
2812 * Actual interfacing with LNet to put/get/register/unregister stuff
2816 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2817 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2819 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2821 struct ptlrpc_bulk_desc *desc;
2824 LASSERT(req != NULL);
2825 desc = req->rq_bulk;
2827 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2828 req->rq_bulk_deadline > cfs_time_current_sec())
2834 spin_lock(&desc->bd_lock);
2835 rc = desc->bd_md_count;
2836 spin_unlock(&desc->bd_lock);
2840 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2841 #define PTLRPC_REPLY_EARLY 0x02
2842 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2843 int ptlrpc_reply(struct ptlrpc_request *req);
2844 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2845 int ptlrpc_error(struct ptlrpc_request *req);
2846 void ptlrpc_resend_req(struct ptlrpc_request *request);
2847 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2848 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2849 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2852 /* ptlrpc/client.c */
2854 * Client-side portals API. Everything to send requests, receive replies,
2855 * request queues, request management, etc.
2858 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2859 struct ptlrpc_client *);
2860 void ptlrpc_cleanup_client(struct obd_import *imp);
2861 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2863 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2864 int ptlrpc_replay_req(struct ptlrpc_request *req);
2865 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2866 void ptlrpc_restart_req(struct ptlrpc_request *req);
2867 void ptlrpc_abort_inflight(struct obd_import *imp);
2868 void ptlrpc_cleanup_imp(struct obd_import *imp);
2869 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2871 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2872 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2874 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2875 set_interpreter_func fn, void *data);
2876 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2877 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2878 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2879 int ptlrpc_expired_set(void *data);
2880 void ptlrpc_interrupted_set(void *data);
2881 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2882 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2883 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2884 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2885 struct ptlrpc_request *req);
2887 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2888 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2890 struct ptlrpc_request_pool *
2891 ptlrpc_init_rq_pool(int, int,
2892 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2894 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2895 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2896 const struct req_format *format);
2897 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2898 struct ptlrpc_request_pool *,
2899 const struct req_format *format);
2900 void ptlrpc_request_free(struct ptlrpc_request *request);
2901 int ptlrpc_request_pack(struct ptlrpc_request *request,
2902 __u32 version, int opcode);
2903 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2904 const struct req_format *format,
2905 __u32 version, int opcode);
2906 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2907 __u32 version, int opcode, char **bufs,
2908 struct ptlrpc_cli_ctx *ctx);
2909 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2910 int opcode, int count, __u32 *lengths,
2912 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2913 __u32 version, int opcode,
2914 int count, __u32 *lengths, char **bufs,
2915 struct ptlrpc_request_pool *pool);
2916 void ptlrpc_req_finished(struct ptlrpc_request *request);
2917 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2918 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2919 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2920 unsigned npages, unsigned max_brw,
2921 unsigned type, unsigned portal);
2922 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2923 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2925 __ptlrpc_free_bulk(bulk, 1);
2927 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2929 __ptlrpc_free_bulk(bulk, 0);
2931 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2932 struct page *page, int pageoffset, int len, int);
2933 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2934 struct page *page, int pageoffset,
2937 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2940 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2941 struct page *page, int pageoffset,
2944 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2947 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2948 struct obd_import *imp);
2949 __u64 ptlrpc_next_xid(void);
2950 __u64 ptlrpc_sample_next_xid(void);
2951 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2953 /* Set of routines to run a function in ptlrpcd context */
2954 void *ptlrpcd_alloc_work(struct obd_import *imp,
2955 int (*cb)(const struct lu_env *, void *), void *data);
2956 void ptlrpcd_destroy_work(void *handler);
2957 int ptlrpcd_queue_work(void *handler);
2960 struct ptlrpc_service_buf_conf {
2961 /* nbufs is buffers # to allocate when growing the pool */
2962 unsigned int bc_nbufs;
2963 /* buffer size to post */
2964 unsigned int bc_buf_size;
2965 /* portal to listed for requests on */
2966 unsigned int bc_req_portal;
2967 /* portal of where to send replies to */
2968 unsigned int bc_rep_portal;
2969 /* maximum request size to be accepted for this service */
2970 unsigned int bc_req_max_size;
2971 /* maximum reply size this service can ever send */
2972 unsigned int bc_rep_max_size;
2975 struct ptlrpc_service_thr_conf {
2976 /* threadname should be 8 characters or less - 6 will be added on */
2978 /* threads increasing factor for each CPU */
2979 unsigned int tc_thr_factor;
2980 /* service threads # to start on each partition while initializing */
2981 unsigned int tc_nthrs_init;
2983 * low water of threads # upper-limit on each partition while running,
2984 * service availability may be impacted if threads number is lower
2985 * than this value. It can be ZERO if the service doesn't require
2986 * CPU affinity or there is only one partition.
2988 unsigned int tc_nthrs_base;
2989 /* "soft" limit for total threads number */
2990 unsigned int tc_nthrs_max;
2991 /* user specified threads number, it will be validated due to
2992 * other members of this structure. */
2993 unsigned int tc_nthrs_user;
2994 /* set NUMA node affinity for service threads */
2995 unsigned int tc_cpu_affinity;
2996 /* Tags for lu_context associated with service thread */
3000 struct ptlrpc_service_cpt_conf {
3001 struct cfs_cpt_table *cc_cptable;
3002 /* string pattern to describe CPTs for a service */
3006 struct ptlrpc_service_conf {
3009 /* soft watchdog timeout multiplifier to print stuck service traces */
3010 unsigned int psc_watchdog_factor;
3011 /* buffer information */
3012 struct ptlrpc_service_buf_conf psc_buf;
3013 /* thread information */
3014 struct ptlrpc_service_thr_conf psc_thr;
3015 /* CPU partition information */
3016 struct ptlrpc_service_cpt_conf psc_cpt;
3017 /* function table */
3018 struct ptlrpc_service_ops psc_ops;
3021 /* ptlrpc/service.c */
3023 * Server-side services API. Register/unregister service, request state
3024 * management, service thread management
3028 void ptlrpc_save_lock(struct ptlrpc_request *req,
3029 struct lustre_handle *lock, int mode, int no_ack);
3030 void ptlrpc_commit_replies(struct obd_export *exp);
3031 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
3032 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
3033 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
3034 struct ptlrpc_service *ptlrpc_register_service(
3035 struct ptlrpc_service_conf *conf,
3036 struct proc_dir_entry *proc_entry);
3037 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
3039 int ptlrpc_start_threads(struct ptlrpc_service *svc);
3040 int ptlrpc_unregister_service(struct ptlrpc_service *service);
3041 int liblustre_check_services(void *arg);
3042 void ptlrpc_daemonize(char *name);
3043 int ptlrpc_service_health_check(struct ptlrpc_service *);
3044 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
3045 void ptlrpc_request_change_export(struct ptlrpc_request *req,
3046 struct obd_export *export);
3048 int ptlrpc_hr_init(void);
3049 void ptlrpc_hr_fini(void);
3053 /* ptlrpc/import.c */
3058 int ptlrpc_connect_import(struct obd_import *imp);
3059 int ptlrpc_init_import(struct obd_import *imp);
3060 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
3061 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
3062 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
3065 /* ptlrpc/pack_generic.c */
3066 int ptlrpc_reconnect_import(struct obd_import *imp);
3070 * ptlrpc msg buffer and swab interface
3074 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
3076 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
3078 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
3079 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
3081 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
3082 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
3084 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
3085 __u32 *lens, char **bufs);
3086 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
3088 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
3089 __u32 *lens, char **bufs, int flags);
3090 #define LPRFL_EARLY_REPLY 1
3091 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
3092 char **bufs, int flags);
3093 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
3094 unsigned int newlen, int move_data);
3095 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
3096 int __lustre_unpack_msg(struct lustre_msg *m, int len);
3097 int lustre_msg_hdr_size(__u32 magic, int count);
3098 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
3099 int lustre_msg_size_v2(int count, __u32 *lengths);
3100 int lustre_packed_msg_size(struct lustre_msg *msg);
3101 int lustre_msg_early_size(void);
3102 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
3103 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
3104 int lustre_msg_buflen(struct lustre_msg *m, int n);
3105 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
3106 int lustre_msg_bufcount(struct lustre_msg *m);
3107 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
3108 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
3109 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
3110 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
3111 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
3112 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
3113 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
3114 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
3115 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
3116 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
3117 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
3118 __u32 lustre_msg_get_type(struct lustre_msg *msg);
3119 __u32 lustre_msg_get_version(struct lustre_msg *msg);
3120 void lustre_msg_add_version(struct lustre_msg *msg, int version);
3121 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
3122 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
3123 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
3124 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
3125 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
3126 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
3127 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
3128 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
3129 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
3130 int lustre_msg_get_status(struct lustre_msg *msg);
3131 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
3132 int lustre_msg_is_v1(struct lustre_msg *msg);
3133 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
3134 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
3135 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
3136 char *lustre_msg_get_jobid(struct lustre_msg *msg);
3137 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
3138 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
3139 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
3141 # warning "remove checksum compatibility support for b1_8"
3142 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
3144 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
3145 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
3146 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
3147 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
3148 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
3149 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
3150 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
3151 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
3152 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
3153 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
3154 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
3155 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
3156 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
3157 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
3158 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
3161 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
3162 unsigned int newlen, int move_data)
3164 LASSERT(req->rq_reply_state);
3165 LASSERT(req->rq_repmsg);
3166 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
3171 /** Change request phase of \a req to \a new_phase */
3173 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
3175 if (req->rq_phase == new_phase)
3178 if (new_phase == RQ_PHASE_UNREGISTERING) {
3179 req->rq_next_phase = req->rq_phase;
3181 atomic_inc(&req->rq_import->imp_unregistering);
3184 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
3186 atomic_dec(&req->rq_import->imp_unregistering);
3189 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
3190 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
3192 req->rq_phase = new_phase;
3196 * Returns true if request \a req got early reply and hard deadline is not met
3199 ptlrpc_client_early(struct ptlrpc_request *req)
3201 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3202 req->rq_reply_deadline > cfs_time_current_sec())
3204 return req->rq_early;
3208 * Returns true if we got real reply from server for this request
3211 ptlrpc_client_replied(struct ptlrpc_request *req)
3213 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3214 req->rq_reply_deadline > cfs_time_current_sec())
3216 return req->rq_replied;
3219 /** Returns true if request \a req is in process of receiving server reply */
3221 ptlrpc_client_recv(struct ptlrpc_request *req)
3223 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3224 req->rq_reply_deadline > cfs_time_current_sec())
3226 return req->rq_receiving_reply;
3230 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3234 spin_lock(&req->rq_lock);
3235 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3236 req->rq_reply_deadline > cfs_time_current_sec()) {
3237 spin_unlock(&req->rq_lock);
3240 rc = req->rq_receiving_reply || req->rq_must_unlink;
3241 spin_unlock(&req->rq_lock);
3246 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3248 if (req->rq_set == NULL)
3249 wake_up(&req->rq_reply_waitq);
3251 wake_up(&req->rq_set->set_waitq);
3255 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3257 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3258 atomic_inc(&rs->rs_refcount);
3262 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3264 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3265 if (atomic_dec_and_test(&rs->rs_refcount))
3266 lustre_free_reply_state(rs);
3269 /* Should only be called once per req */
3270 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3272 if (req->rq_reply_state == NULL)
3273 return; /* shouldn't occur */
3274 ptlrpc_rs_decref(req->rq_reply_state);
3275 req->rq_reply_state = NULL;
3276 req->rq_repmsg = NULL;
3279 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3281 return lustre_msg_get_magic(req->rq_reqmsg);
3284 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3286 switch (req->rq_reqmsg->lm_magic) {
3287 case LUSTRE_MSG_MAGIC_V2:
3288 return req->rq_reqmsg->lm_repsize;
3290 LASSERTF(0, "incorrect message magic: %08x\n",
3291 req->rq_reqmsg->lm_magic);
3296 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3298 if (req->rq_delay_limit != 0 &&
3299 cfs_time_before(cfs_time_add(req->rq_queued_time,
3300 cfs_time_seconds(req->rq_delay_limit)),
3301 cfs_time_current())) {
3307 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3309 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3310 spin_lock(&req->rq_lock);
3311 req->rq_no_resend = 1;
3312 spin_unlock(&req->rq_lock);
3314 return req->rq_no_resend;
3318 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3320 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3322 return svcpt->scp_service->srv_watchdog_factor *
3323 max_t(int, at, obd_timeout);
3326 static inline struct ptlrpc_service *
3327 ptlrpc_req2svc(struct ptlrpc_request *req)
3329 LASSERT(req->rq_rqbd != NULL);
3330 return req->rq_rqbd->rqbd_svcpt->scp_service;
3333 /* ldlm/ldlm_lib.c */
3335 * Target client logic
3338 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3339 int client_obd_cleanup(struct obd_device *obddev);
3340 int client_connect_import(const struct lu_env *env,
3341 struct obd_export **exp, struct obd_device *obd,
3342 struct obd_uuid *cluuid, struct obd_connect_data *,
3344 int client_disconnect_export(struct obd_export *exp);
3345 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3347 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3348 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3349 struct obd_uuid *uuid);
3350 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3351 void client_destroy_import(struct obd_import *imp);
3355 /* ptlrpc/pinger.c */
3357 * Pinger API (client side only)
3360 enum timeout_event {
3363 struct timeout_item;
3364 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3365 int ptlrpc_pinger_add_import(struct obd_import *imp);
3366 int ptlrpc_pinger_del_import(struct obd_import *imp);
3367 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3368 timeout_cb_t cb, void *data,
3369 struct list_head *obd_list);
3370 int ptlrpc_del_timeout_client(struct list_head *obd_list,
3371 enum timeout_event event);
3372 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3373 int ptlrpc_obd_ping(struct obd_device *obd);
3374 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3375 void ping_evictor_start(void);
3376 void ping_evictor_stop(void);
3377 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3378 void ptlrpc_pinger_ir_up(void);
3379 void ptlrpc_pinger_ir_down(void);
3381 int ptlrpc_pinger_suppress_pings(void);
3383 /* ptlrpc daemon bind policy */
3385 /* all ptlrpcd threads are free mode */
3386 PDB_POLICY_NONE = 1,
3387 /* all ptlrpcd threads are bound mode */
3388 PDB_POLICY_FULL = 2,
3389 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3390 PDB_POLICY_PAIR = 3,
3391 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3392 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3393 * If kernel supports NUMA, pthrpcd threads are binded and
3394 * grouped by NUMA node */
3395 PDB_POLICY_NEIGHBOR = 4,
3398 /* ptlrpc daemon load policy
3399 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3400 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3401 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3402 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3403 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3405 /* on the same CPU core as the caller */
3406 PDL_POLICY_SAME = 1,
3407 /* within the same CPU partition, but not the same core as the caller */
3408 PDL_POLICY_LOCAL = 2,
3409 /* round-robin on all CPU cores, but not the same core as the caller */
3410 PDL_POLICY_ROUND = 3,
3411 /* the specified CPU core is preferred, but not enforced */
3412 PDL_POLICY_PREFERRED = 4,
3415 /* ptlrpc/ptlrpcd.c */
3416 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3417 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3418 void ptlrpcd_wake(struct ptlrpc_request *req);
3419 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3420 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3421 int ptlrpcd_addref(void);
3422 void ptlrpcd_decref(void);
3424 /* ptlrpc/lproc_ptlrpc.c */
3426 * procfs output related functions
3429 const char* ll_opcode2str(__u32 opcode);
3431 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3432 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3433 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3435 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3436 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3437 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3441 /* ptlrpc/llog_server.c */
3442 int llog_origin_handle_open(struct ptlrpc_request *req);
3443 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3444 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3445 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3446 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3447 int llog_origin_handle_close(struct ptlrpc_request *req);
3448 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3450 /* ptlrpc/llog_client.c */
3451 extern struct llog_operations llog_client_ops;