4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #include <linux/lustre_net.h>
59 #include <libcfs/libcfs.h>
60 #include <lnet/lnet.h>
61 #include <lustre/lustre_idl.h>
62 #include <lustre_ha.h>
63 #include <lustre_sec.h>
64 #include <lustre_import.h>
65 #include <lprocfs_status.h>
66 #include <lu_object.h>
67 #include <lustre_req_layout.h>
68 #include <obd_support.h>
69 #include <lustre_ver.h>
71 /* MD flags we _always_ use */
72 #define PTLRPC_MD_OPTIONS 0
75 * Max # of bulk operations in one request.
76 * In order for the client and server to properly negotiate the maximum
77 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
78 * value. The client is free to limit the actual RPC size for any bulk
79 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
80 #define PTLRPC_BULK_OPS_BITS 2
81 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
83 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
84 * should not be used on the server at all. Otherwise, it imposes a
85 * protocol limitation on the maximum RPC size that can be used by any
86 * RPC sent to that server in the future. Instead, the server should
87 * use the negotiated per-client ocd_brw_size to determine the bulk
89 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
92 * Define maxima for bulk I/O.
94 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
95 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
96 * currently supported maximum between peers at connect via ocd_brw_size.
98 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
99 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
100 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
102 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
103 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
104 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
105 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
106 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
107 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
109 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
110 #if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
111 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
113 #if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
114 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
116 #if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
117 # error "PTLRPC_MAX_BRW_SIZE too big"
119 #if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_PAGES too big"
123 #define PTLRPC_NTHRS_INIT 2
128 * Constants determine how memory is used to buffer incoming service requests.
130 * ?_NBUFS # buffers to allocate when growing the pool
131 * ?_BUFSIZE # bytes in a single request buffer
132 * ?_MAXREQSIZE # maximum request service will receive
134 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
135 * of ?_NBUFS is added to the pool.
137 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
138 * considered full when less than ?_MAXREQSIZE is left in them.
143 * Constants determine how threads are created for ptlrpc service.
145 * ?_NTHRS_INIT # threads to create for each service partition on
146 * initializing. If it's non-affinity service and
147 * there is only one partition, it's the overall #
148 * threads for the service while initializing.
149 * ?_NTHRS_BASE # threads should be created at least for each
150 * ptlrpc partition to keep the service healthy.
151 * It's the low-water mark of threads upper-limit
152 * for each partition.
153 * ?_THR_FACTOR # threads can be added on threads upper-limit for
154 * each CPU core. This factor is only for reference,
155 * we might decrease value of factor if number of cores
156 * per CPT is above a limit.
157 * ?_NTHRS_MAX # overall threads can be created for a service,
158 * it's a soft limit because if service is running
159 * on machine with hundreds of cores and tens of
160 * CPU partitions, we need to guarantee each partition
161 * has ?_NTHRS_BASE threads, which means total threads
162 * will be ?_NTHRS_BASE * number_of_cpts which can
163 * exceed ?_NTHRS_MAX.
167 * #define MDS_NTHRS_INIT 2
168 * #define MDS_NTHRS_BASE 64
169 * #define MDS_NTHRS_FACTOR 8
170 * #define MDS_NTHRS_MAX 1024
173 * ---------------------------------------------------------------------
174 * Server(A) has 16 cores, user configured it to 4 partitions so each
175 * partition has 4 cores, then actual number of service threads on each
177 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
179 * Total number of threads for the service is:
180 * 96 * partitions(4) = 384
183 * ---------------------------------------------------------------------
184 * Server(B) has 32 cores, user configured it to 4 partitions so each
185 * partition has 8 cores, then actual number of service threads on each
187 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
189 * Total number of threads for the service is:
190 * 128 * partitions(4) = 512
193 * ---------------------------------------------------------------------
194 * Server(B) has 96 cores, user configured it to 8 partitions so each
195 * partition has 12 cores, then actual number of service threads on each
197 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
199 * Total number of threads for the service is:
200 * 160 * partitions(8) = 1280
202 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
203 * as upper limit of threads number for each partition:
204 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
207 * ---------------------------------------------------------------------
208 * Server(C) have a thousand of cores and user configured it to 32 partitions
209 * MDS_NTHRS_BASE(64) * 32 = 2048
211 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
212 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
213 * to keep service healthy, so total number of threads will just be 2048.
215 * NB: we don't suggest to choose server with that many cores because backend
216 * filesystem itself, buffer cache, or underlying network stack might
217 * have some SMP scalability issues at that large scale.
219 * If user already has a fat machine with hundreds or thousands of cores,
220 * there are two choices for configuration:
221 * a) create CPU table from subset of all CPUs and run Lustre on
223 * b) bind service threads on a few partitions, see modparameters of
224 * MDS and OSS for details
226 * NB: these calculations (and examples below) are simplified to help
227 * understanding, the real implementation is a little more complex,
228 * please see ptlrpc_server_nthreads_check() for details.
233 * LDLM threads constants:
235 * Given 8 as factor and 24 as base threads number
238 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
241 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
242 * threads for each partition and total threads number will be 112.
245 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
246 * threads for each partition to keep service healthy, so total threads
247 * number should be 24 * 8 = 192.
249 * So with these constants, threads number will be at the similar level
250 * of old versions, unless target machine has over a hundred cores
252 #define LDLM_THR_FACTOR 8
253 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
254 #define LDLM_NTHRS_BASE 24
255 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
257 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
258 #define LDLM_CLIENT_NBUFS 1
259 #define LDLM_SERVER_NBUFS 64
260 #define LDLM_BUFSIZE (8 * 1024)
261 #define LDLM_MAXREQSIZE (5 * 1024)
262 #define LDLM_MAXREPSIZE (1024)
265 * MDS threads constants:
267 * Please see examples in "Thread Constants", MDS threads number will be at
268 * the comparable level of old versions, unless the server has many cores.
270 #ifndef MDS_MAX_THREADS
271 #define MDS_MAX_THREADS 1024
272 #define MDS_MAX_OTHR_THREADS 256
274 #else /* MDS_MAX_THREADS */
275 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
276 #undef MDS_MAX_THREADS
277 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
279 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
282 /* default service */
283 #define MDS_THR_FACTOR 8
284 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
285 #define MDS_NTHRS_MAX MDS_MAX_THREADS
286 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
288 /* read-page service */
289 #define MDS_RDPG_THR_FACTOR 4
290 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
291 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
292 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
294 /* these should be removed when we remove setattr service in the future */
295 #define MDS_SETA_THR_FACTOR 4
296 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
297 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
298 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
300 /* non-affinity threads */
301 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
302 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
307 * Assume file name length = FNAME_MAX = 256 (true for ext3).
308 * path name length = PATH_MAX = 4096
309 * LOV MD size max = EA_MAX = 24 * 2000
310 * (NB: 24 is size of lov_ost_data)
311 * LOV LOGCOOKIE size max = 32 * 2000
312 * (NB: 32 is size of llog_cookie)
313 * symlink: FNAME_MAX + PATH_MAX <- largest
314 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
315 * rename: FNAME_MAX + FNAME_MAX
316 * open: FNAME_MAX + EA_MAX
318 * MDS_MAXREQSIZE ~= 4736 bytes =
319 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
320 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
322 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
323 * except in the open case where there are a large number of OSTs in a LOV.
325 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
326 #define MDS_MAXREPSIZE (9 * 1024) /* >= 8300 */
329 * MDS incoming request with LOV EA
330 * 24 = sizeof(struct lov_ost_data), i.e: replay of opencreate
332 #define MDS_LOV_MAXREQSIZE max(MDS_MAXREQSIZE, \
333 362 + LOV_MAX_STRIPE_COUNT * 24)
335 * MDS outgoing reply with LOV EA
337 * NB: max reply size Lustre 2.4+ client can get from old MDS is:
338 * LOV_MAX_STRIPE_COUNT * (llog_cookie + lov_ost_data) + extra bytes
340 * but 2.4 or later MDS will never send reply with llog_cookie to any
341 * version client. This macro is defined for server side reply buffer size.
343 #define MDS_LOV_MAXREPSIZE MDS_LOV_MAXREQSIZE
346 * This is the size of a maximum REINT_SETXATTR request:
348 * lustre_msg 56 (32 + 4 x 5 + 4)
350 * mdt_rec_setxattr 136
352 * name 256 (XATTR_NAME_MAX)
353 * value 65536 (XATTR_SIZE_MAX)
355 #define MDS_EA_MAXREQSIZE 66288
358 * These are the maximum request and reply sizes (rounded up to 1 KB
359 * boundaries) for the "regular" MDS_REQUEST_PORTAL and MDS_REPLY_PORTAL.
361 #define MDS_REG_MAXREQSIZE (((max(MDS_EA_MAXREQSIZE, \
362 MDS_LOV_MAXREQSIZE) + 1023) >> 10) << 10)
363 #define MDS_REG_MAXREPSIZE MDS_REG_MAXREQSIZE
366 * The update request includes all of updates from the create, which might
367 * include linkea (4K maxim), together with other updates, we set it to 9K:
368 * lustre_msg + ptlrpc_body + UPDATE_BUF_SIZE (8K)
370 #define OUT_MAXREQSIZE (9 * 1024)
371 #define OUT_MAXREPSIZE MDS_MAXREPSIZE
373 /** MDS_BUFSIZE = max_reqsize (w/o LOV EA) + max sptlrpc payload size */
374 #define MDS_BUFSIZE max(MDS_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
378 * MDS_REG_BUFSIZE should at least be MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD.
379 * However, we need to allocate a much larger buffer for it because LNet
380 * requires each MD(rqbd) has at least MDS_REQ_MAXREQSIZE bytes left to avoid
381 * dropping of maximum-sized incoming request. So if MDS_REG_BUFSIZE is only a
382 * little larger than MDS_REG_MAXREQSIZE, then it can only fit in one request
383 * even there are about MDS_REG_MAX_REQSIZE bytes left in a rqbd, and memory
384 * utilization is very low.
386 * In the meanwhile, size of rqbd can't be too large, because rqbd can't be
387 * reused until all requests fit in it have been processed and released,
388 * which means one long blocked request can prevent the rqbd be reused.
389 * Now we set request buffer size to 160 KB, so even each rqbd is unlinked
390 * from LNet with unused 65 KB, buffer utilization will be about 59%.
391 * Please check LU-2432 for details.
393 #define MDS_REG_BUFSIZE max(MDS_REG_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
397 * OUT_BUFSIZE = max_out_reqsize + max sptlrpc payload (~1K) which is
398 * about 10K, for the same reason as MDS_REG_BUFSIZE, we also give some
399 * extra bytes to each request buffer to improve buffer utilization rate.
401 #define OUT_BUFSIZE max(OUT_MAXREQSIZE + SPTLRPC_MAX_PAYLOAD, \
404 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
405 #define FLD_MAXREQSIZE (160)
407 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
408 #define FLD_MAXREPSIZE (152)
409 #define FLD_BUFSIZE (1 << 12)
412 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
414 #define SEQ_MAXREQSIZE (160)
416 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
417 #define SEQ_MAXREPSIZE (152)
418 #define SEQ_BUFSIZE (1 << 12)
420 /** MGS threads must be >= 3, see bug 22458 comment #28 */
421 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
422 #define MGS_NTHRS_MAX 32
425 #define MGS_BUFSIZE (8 * 1024)
426 #define MGS_MAXREQSIZE (7 * 1024)
427 #define MGS_MAXREPSIZE (9 * 1024)
430 * OSS threads constants:
432 * Given 8 as factor and 64 as base threads number
435 * On 8-core server configured to 2 partitions, we will have
436 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
439 * On 32-core machine configured to 4 partitions, we will have
440 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
441 * will be 112 * 4 = 448.
444 * On 64-core machine configured to 4 partitions, we will have
445 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
446 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
447 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
448 * for each partition.
450 * So we can see that with these constants, threads number wil be at the
451 * similar level of old versions, unless the server has many cores.
453 /* depress threads factor for VM with small memory size */
454 #define OSS_THR_FACTOR min_t(int, 8, \
455 NUM_CACHEPAGES >> (28 - PAGE_CACHE_SHIFT))
456 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
457 #define OSS_NTHRS_BASE 64
458 #define OSS_NTHRS_MAX 512
460 /* threads for handling "create" request */
461 #define OSS_CR_THR_FACTOR 1
462 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
463 #define OSS_CR_NTHRS_BASE 8
464 #define OSS_CR_NTHRS_MAX 64
467 * OST_IO_MAXREQSIZE ~=
468 * lustre_msg + ptlrpc_body + obdo + obd_ioobj +
469 * DT_MAX_BRW_PAGES * niobuf_remote
471 * - single object with 16 pages is 512 bytes
472 * - OST_IO_MAXREQSIZE must be at least 1 page of cookies plus some spillover
473 * - Must be a multiple of 1024
474 * - actual size is about 18K
476 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + \
477 sizeof(struct ptlrpc_body) + \
478 sizeof(struct obdo) + \
479 sizeof(struct obd_ioobj) + \
480 sizeof(struct niobuf_remote) * DT_MAX_BRW_PAGES)
482 * FIEMAP request can be 4K+ for now
484 #define OST_MAXREQSIZE (16 * 1024)
485 #define OST_IO_MAXREQSIZE max_t(int, OST_MAXREQSIZE, \
486 (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1))
488 #define OST_MAXREPSIZE (9 * 1024)
489 #define OST_IO_MAXREPSIZE OST_MAXREPSIZE
492 /** OST_BUFSIZE = max_reqsize + max sptlrpc payload size */
493 #define OST_BUFSIZE max_t(int, OST_MAXREQSIZE + 1024, 16 * 1024)
495 * OST_IO_MAXREQSIZE is 18K, giving extra 46K can increase buffer utilization
496 * rate of request buffer, please check comment of MDS_LOV_BUFSIZE for details.
498 #define OST_IO_BUFSIZE max_t(int, OST_IO_MAXREQSIZE + 1024, 64 * 1024)
500 /* Macro to hide a typecast. */
501 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
504 * Structure to single define portal connection.
506 struct ptlrpc_connection {
507 /** linkage for connections hash table */
508 struct hlist_node c_hash;
509 /** Our own lnet nid for this connection */
511 /** Remote side nid for this connection */
512 lnet_process_id_t c_peer;
513 /** UUID of the other side */
514 struct obd_uuid c_remote_uuid;
515 /** reference counter for this connection */
519 /** Client definition for PortalRPC */
520 struct ptlrpc_client {
521 /** What lnet portal does this client send messages to by default */
522 __u32 cli_request_portal;
523 /** What portal do we expect replies on */
524 __u32 cli_reply_portal;
525 /** Name of the client */
529 /** state flags of requests */
530 /* XXX only ones left are those used by the bulk descs as well! */
531 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
532 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
534 #define REQ_MAX_ACK_LOCKS 8
536 union ptlrpc_async_args {
538 * Scratchpad for passing args to completion interpreter. Users
539 * cast to the struct of their choosing, and CLASSERT that this is
540 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
541 * a pointer to it here. The pointer_arg ensures this struct is at
542 * least big enough for that.
544 void *pointer_arg[11];
548 struct ptlrpc_request_set;
549 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
550 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
553 * Definition of request set structure.
554 * Request set is a list of requests (not necessary to the same target) that
555 * once populated with RPCs could be sent in parallel.
556 * There are two kinds of request sets. General purpose and with dedicated
557 * serving thread. Example of the latter is ptlrpcd set.
558 * For general purpose sets once request set started sending it is impossible
559 * to add new requests to such set.
560 * Provides a way to call "completion callbacks" when all requests in the set
563 struct ptlrpc_request_set {
564 atomic_t set_refcount;
565 /** number of in queue requests */
566 atomic_t set_new_count;
567 /** number of uncompleted requests */
568 atomic_t set_remaining;
569 /** wait queue to wait on for request events */
570 wait_queue_head_t set_waitq;
571 wait_queue_head_t *set_wakeup_ptr;
572 /** List of requests in the set */
573 struct list_head set_requests;
575 * List of completion callbacks to be called when the set is completed
576 * This is only used if \a set_interpret is NULL.
577 * Links struct ptlrpc_set_cbdata.
579 struct list_head set_cblist;
580 /** Completion callback, if only one. */
581 set_interpreter_func set_interpret;
582 /** opaq argument passed to completion \a set_interpret callback. */
585 * Lock for \a set_new_requests manipulations
586 * locked so that any old caller can communicate requests to
587 * the set holder who can then fold them into the lock-free set
589 spinlock_t set_new_req_lock;
590 /** List of new yet unsent requests. Only used with ptlrpcd now. */
591 struct list_head set_new_requests;
593 /** rq_status of requests that have been freed already */
595 /** Additional fields used by the flow control extension */
596 /** Maximum number of RPCs in flight */
597 int set_max_inflight;
598 /** Callback function used to generate RPCs */
599 set_producer_func set_producer;
600 /** opaq argument passed to the producer callback */
601 void *set_producer_arg;
605 * Description of a single ptrlrpc_set callback
607 struct ptlrpc_set_cbdata {
608 /** List linkage item */
609 struct list_head psc_item;
610 /** Pointer to interpreting function */
611 set_interpreter_func psc_interpret;
612 /** Opaq argument to pass to the callback */
616 struct ptlrpc_bulk_desc;
617 struct ptlrpc_service_part;
618 struct ptlrpc_service;
621 * ptlrpc callback & work item stuff
623 struct ptlrpc_cb_id {
624 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
625 void *cbid_arg; /* additional arg */
628 /** Maximum number of locks to fit into reply state */
629 #define RS_MAX_LOCKS 8
633 * Structure to define reply state on the server
634 * Reply state holds various reply message information. Also for "difficult"
635 * replies (rep-ack case) we store the state after sending reply and wait
636 * for the client to acknowledge the reception. In these cases locks could be
637 * added to the state for replay/failover consistency guarantees.
639 struct ptlrpc_reply_state {
640 /** Callback description */
641 struct ptlrpc_cb_id rs_cb_id;
642 /** Linkage for list of all reply states in a system */
643 struct list_head rs_list;
644 /** Linkage for list of all reply states on same export */
645 struct list_head rs_exp_list;
646 /** Linkage for list of all reply states for same obd */
647 struct list_head rs_obd_list;
649 struct list_head rs_debug_list;
651 /** A spinlock to protect the reply state flags */
653 /** Reply state flags */
654 unsigned long rs_difficult:1; /* ACK/commit stuff */
655 unsigned long rs_no_ack:1; /* no ACK, even for
656 difficult requests */
657 unsigned long rs_scheduled:1; /* being handled? */
658 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
659 unsigned long rs_handled:1; /* been handled yet? */
660 unsigned long rs_on_net:1; /* reply_out_callback pending? */
661 unsigned long rs_prealloc:1; /* rs from prealloc list */
662 unsigned long rs_committed:1;/* the transaction was committed
663 and the rs was dispatched
664 by ptlrpc_commit_replies */
665 /** Size of the state */
669 /** Transaction number */
673 struct obd_export *rs_export;
674 struct ptlrpc_service_part *rs_svcpt;
675 /** Lnet metadata handle for the reply */
676 lnet_handle_md_t rs_md_h;
677 atomic_t rs_refcount;
679 /** Context for the sevice thread */
680 struct ptlrpc_svc_ctx *rs_svc_ctx;
681 /** Reply buffer (actually sent to the client), encoded if needed */
682 struct lustre_msg *rs_repbuf; /* wrapper */
683 /** Size of the reply buffer */
684 int rs_repbuf_len; /* wrapper buf length */
685 /** Size of the reply message */
686 int rs_repdata_len; /* wrapper msg length */
688 * Actual reply message. Its content is encrupted (if needed) to
689 * produce reply buffer for actual sending. In simple case
690 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
692 struct lustre_msg *rs_msg; /* reply message */
694 /** Number of locks awaiting client ACK */
696 /** Handles of locks awaiting client reply ACK */
697 struct lustre_handle rs_locks[RS_MAX_LOCKS];
698 /** Lock modes of locks in \a rs_locks */
699 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
702 struct ptlrpc_thread;
706 RQ_PHASE_NEW = 0xebc0de00,
707 RQ_PHASE_RPC = 0xebc0de01,
708 RQ_PHASE_BULK = 0xebc0de02,
709 RQ_PHASE_INTERPRET = 0xebc0de03,
710 RQ_PHASE_COMPLETE = 0xebc0de04,
711 RQ_PHASE_UNREGISTERING = 0xebc0de05,
712 RQ_PHASE_UNDEFINED = 0xebc0de06
715 /** Type of request interpreter call-back */
716 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
717 struct ptlrpc_request *req,
719 /** Type of request resend call-back */
720 typedef void (*ptlrpc_resend_cb_t)(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
815 * \param[in,out] arg A generic char buffer
817 * \see nrs_policy_start_locked()
819 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy,
822 * Called when deactivating a policy via lprocfs; policies deallocate
823 * their resources here; this operation is optional
825 * \param[in,out] policy The policy being stopped
827 * \see nrs_policy_stop0()
829 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
831 * Used for policy-specific operations; i.e. not generic ones like
832 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
833 * to an ioctl; this operation is optional.
835 * \param[in,out] policy The policy carrying out operation \a opc
836 * \param[in] opc The command operation being carried out
837 * \param[in,out] arg An generic buffer for communication between the
838 * user and the control operation
843 * \see ptlrpc_nrs_policy_control()
845 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
846 enum ptlrpc_nrs_ctl opc, void *arg);
849 * Called when obtaining references to the resources of the resource
850 * hierarchy for a request that has arrived for handling at the PTLRPC
851 * service. Policies should return -ve for requests they do not wish
852 * to handle. This operation is mandatory.
854 * \param[in,out] policy The policy we're getting resources for.
855 * \param[in,out] nrq The request we are getting resources for.
856 * \param[in] parent The parent resource of the resource being
857 * requested; set to NULL if none.
858 * \param[out] resp The resource is to be returned here; the
859 * fallback policy in an NRS head should
860 * \e always return a non-NULL pointer value.
861 * \param[in] moving_req When set, signifies that this is an attempt
862 * to obtain resources for a request being moved
863 * to the high-priority NRS head by
864 * ldlm_lock_reorder_req().
865 * This implies two things:
866 * 1. We are under obd_export::exp_rpc_lock and
867 * so should not sleep.
868 * 2. We should not perform non-idempotent or can
869 * skip performing idempotent operations that
870 * were carried out when resources were first
871 * taken for the request when it was initialized
872 * in ptlrpc_nrs_req_initialize().
874 * \retval 0, +ve The level of the returned resource in the resource
875 * hierarchy; currently only 0 (for a non-leaf resource)
876 * and 1 (for a leaf resource) are supported by the
880 * \see ptlrpc_nrs_req_initialize()
881 * \see ptlrpc_nrs_hpreq_add_nolock()
882 * \see ptlrpc_nrs_req_hp_move()
884 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
885 struct ptlrpc_nrs_request *nrq,
886 const struct ptlrpc_nrs_resource *parent,
887 struct ptlrpc_nrs_resource **resp,
890 * Called when releasing references taken for resources in the resource
891 * hierarchy for the request; this operation is optional.
893 * \param[in,out] policy The policy the resource belongs to
894 * \param[in] res The resource to be freed
896 * \see ptlrpc_nrs_req_finalize()
897 * \see ptlrpc_nrs_hpreq_add_nolock()
898 * \see ptlrpc_nrs_req_hp_move()
900 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
901 const struct ptlrpc_nrs_resource *res);
904 * Obtains a request for handling from the policy, and optionally
905 * removes the request from the policy; this operation is mandatory.
907 * \param[in,out] policy The policy to poll
908 * \param[in] peek When set, signifies that we just want to
909 * examine the request, and not handle it, so the
910 * request is not removed from the policy.
911 * \param[in] force When set, it will force a policy to return a
912 * request if it has one queued.
914 * \retval NULL No request available for handling
915 * \retval valid-pointer The request polled for handling
917 * \see ptlrpc_nrs_req_get_nolock()
919 struct ptlrpc_nrs_request *
920 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
923 * Called when attempting to add a request to a policy for later
924 * handling; this operation is mandatory.
926 * \param[in,out] policy The policy on which to enqueue \a nrq
927 * \param[in,out] nrq The request to enqueue
932 * \see ptlrpc_nrs_req_add_nolock()
934 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
935 struct ptlrpc_nrs_request *nrq);
937 * Removes a request from the policy's set of pending requests. Normally
938 * called after a request has been polled successfully from the policy
939 * for handling; this operation is mandatory.
941 * \param[in,out] policy The policy the request \a nrq belongs to
942 * \param[in,out] nrq The request to dequeue
944 * \see ptlrpc_nrs_req_del_nolock()
946 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
947 struct ptlrpc_nrs_request *nrq);
949 * Called after the request being carried out. Could be used for
950 * job/resource control; this operation is optional.
952 * \param[in,out] policy The policy which is stopping to handle request
954 * \param[in,out] nrq The request
956 * \pre assert_spin_locked(&svcpt->scp_req_lock)
958 * \see ptlrpc_nrs_req_stop_nolock()
960 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
961 struct ptlrpc_nrs_request *nrq);
963 * Registers the policy's lprocfs interface with a PTLRPC service.
965 * \param[in] svc The service
970 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
972 * Unegisters the policy's lprocfs interface with a PTLRPC service.
974 * In cases of failed policy registration in
975 * \e ptlrpc_nrs_policy_register(), this function may be called for a
976 * service which has not registered the policy successfully, so
977 * implementations of this method should make sure their operations are
978 * safe in such cases.
980 * \param[in] svc The service
982 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
988 enum nrs_policy_flags {
990 * Fallback policy, use this flag only on a single supported policy per
991 * service. The flag cannot be used on policies that use
992 * \e PTLRPC_NRS_FL_REG_EXTERN
994 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
996 * Start policy immediately after registering.
998 PTLRPC_NRS_FL_REG_START = (1 << 1),
1000 * This is a policy registering from a module different to the one NRS
1001 * core ships in (currently ptlrpc).
1003 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
1009 * Denotes whether an NRS instance is for handling normal or high-priority
1010 * RPCs, or whether an operation pertains to one or both of the NRS instances
1013 enum ptlrpc_nrs_queue_type {
1014 PTLRPC_NRS_QUEUE_REG = (1 << 0),
1015 PTLRPC_NRS_QUEUE_HP = (1 << 1),
1016 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
1022 * A PTLRPC service has at least one NRS head instance for handling normal
1023 * priority RPCs, and may optionally have a second NRS head instance for
1024 * handling high-priority RPCs. Each NRS head maintains a list of available
1025 * policies, of which one and only one policy is acting as the fallback policy,
1026 * and optionally a different policy may be acting as the primary policy. For
1027 * all RPCs handled by this NRS head instance, NRS core will first attempt to
1028 * enqueue the RPC using the primary policy (if any). The fallback policy is
1029 * used in the following cases:
1030 * - when there was no primary policy in the
1031 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
1033 * - when the primary policy that was at the
1034 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1035 * RPC was initialized, denoted it did not wish, or for some other reason was
1036 * not able to handle the request, by returning a non-valid NRS resource
1038 * - when the primary policy that was at the
1039 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
1040 * RPC was initialized, fails later during the request enqueueing stage.
1042 * \see nrs_resource_get_safe()
1043 * \see nrs_request_enqueue()
1046 spinlock_t nrs_lock;
1047 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
1049 * List of registered policies
1051 struct list_head nrs_policy_list;
1053 * List of policies with queued requests. Policies that have any
1054 * outstanding requests are queued here, and this list is queried
1055 * in a round-robin manner from NRS core when obtaining a request
1056 * for handling. This ensures that requests from policies that at some
1057 * point transition away from the
1058 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
1060 struct list_head nrs_policy_queued;
1062 * Service partition for this NRS head
1064 struct ptlrpc_service_part *nrs_svcpt;
1066 * Primary policy, which is the preferred policy for handling RPCs
1068 struct ptlrpc_nrs_policy *nrs_policy_primary;
1070 * Fallback policy, which is the backup policy for handling RPCs
1072 struct ptlrpc_nrs_policy *nrs_policy_fallback;
1074 * This NRS head handles either HP or regular requests
1076 enum ptlrpc_nrs_queue_type nrs_queue_type;
1078 * # queued requests from all policies in this NRS head
1080 unsigned long nrs_req_queued;
1082 * # scheduled requests from all policies in this NRS head
1084 unsigned long nrs_req_started;
1086 * # policies on this NRS
1088 unsigned nrs_num_pols;
1090 * This NRS head is in progress of starting a policy
1092 unsigned nrs_policy_starting:1;
1094 * In progress of shutting down the whole NRS head; used during
1097 unsigned nrs_stopping:1;
1099 * NRS policy is throttling reqeust
1101 unsigned nrs_throttling:1;
1104 #define NRS_POL_NAME_MAX 16
1106 struct ptlrpc_nrs_pol_desc;
1109 * Service compatibility predicate; this determines whether a policy is adequate
1110 * for handling RPCs of a particular PTLRPC service.
1112 * XXX:This should give the same result during policy registration and
1113 * unregistration, and for all partitions of a service; so the result should not
1114 * depend on temporal service or other properties, that may influence the
1117 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
1118 const struct ptlrpc_nrs_pol_desc *desc);
1120 struct ptlrpc_nrs_pol_conf {
1122 * Human-readable policy name
1124 char nc_name[NRS_POL_NAME_MAX];
1126 * NRS operations for this policy
1128 const struct ptlrpc_nrs_pol_ops *nc_ops;
1130 * Service compatibility predicate
1132 nrs_pol_desc_compat_t nc_compat;
1134 * Set for policies that support a single ptlrpc service, i.e. ones that
1135 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
1136 * depicts the name of the single service that such policies are
1139 const char *nc_compat_svc_name;
1141 * Owner module for this policy descriptor; policies registering from a
1142 * different module to the one the NRS framework is held within
1143 * (currently ptlrpc), should set this field to THIS_MODULE.
1145 struct module *nc_owner;
1147 * Policy registration flags; a bitmast of \e nrs_policy_flags
1153 * NRS policy registering descriptor
1155 * Is used to hold a description of a policy that can be passed to NRS core in
1156 * order to register the policy with NRS heads in different PTLRPC services.
1158 struct ptlrpc_nrs_pol_desc {
1160 * Human-readable policy name
1162 char pd_name[NRS_POL_NAME_MAX];
1164 * Link into nrs_core::nrs_policies
1166 struct list_head pd_list;
1168 * NRS operations for this policy
1170 const struct ptlrpc_nrs_pol_ops *pd_ops;
1172 * Service compatibility predicate
1174 nrs_pol_desc_compat_t pd_compat;
1176 * Set for policies that are compatible with only one PTLRPC service.
1178 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
1180 const char *pd_compat_svc_name;
1182 * Owner module for this policy descriptor.
1184 * We need to hold a reference to the module whenever we might make use
1185 * of any of the module's contents, i.e.
1186 * - If one or more instances of the policy are at a state where they
1187 * might be handling a request, i.e.
1188 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1189 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
1190 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
1191 * is taken on the module when
1192 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
1193 * becomes 0, so that we hold only one reference to the module maximum
1196 * We do not need to hold a reference to the module, even though we
1197 * might use code and data from the module, in the following cases:
1198 * - During external policy registration, because this should happen in
1199 * the module's init() function, in which case the module is safe from
1200 * removal because a reference is being held on the module by the
1201 * kernel, and iirc kmod (and I guess module-init-tools also) will
1202 * serialize any racing processes properly anyway.
1203 * - During external policy unregistration, because this should happen
1204 * in a module's exit() function, and any attempts to start a policy
1205 * instance would need to take a reference on the module, and this is
1206 * not possible once we have reached the point where the exit()
1207 * handler is called.
1208 * - During service registration and unregistration, as service setup
1209 * and cleanup, and policy registration, unregistration and policy
1210 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
1211 * as long as users adhere to the convention of registering policies
1212 * in init() and unregistering them in module exit() functions, there
1213 * should not be a race between these operations.
1214 * - During any policy-specific lprocfs operations, because a reference
1215 * is held by the kernel on a proc entry that has been entered by a
1216 * syscall, so as long as proc entries are removed during unregistration time,
1217 * then unregistration and lprocfs operations will be properly
1220 struct module *pd_owner;
1222 * Bitmask of \e nrs_policy_flags
1226 * # of references on this descriptor
1234 * Policies transition from one state to the other during their lifetime
1236 enum ptlrpc_nrs_pol_state {
1238 * Not a valid policy state.
1240 NRS_POL_STATE_INVALID,
1242 * Policies are at this state either at the start of their life, or
1243 * transition here when the user selects a different policy to act
1244 * as the primary one.
1246 NRS_POL_STATE_STOPPED,
1248 * Policy is progress of stopping
1250 NRS_POL_STATE_STOPPING,
1252 * Policy is in progress of starting
1254 NRS_POL_STATE_STARTING,
1256 * A policy is in this state in two cases:
1257 * - it is the fallback policy, which is always in this state.
1258 * - it has been activated by the user; i.e. it is the primary policy,
1260 NRS_POL_STATE_STARTED,
1264 * NRS policy information
1266 * Used for obtaining information for the status of a policy via lprocfs
1268 struct ptlrpc_nrs_pol_info {
1272 char pi_name[NRS_POL_NAME_MAX];
1274 * Current policy state
1276 enum ptlrpc_nrs_pol_state pi_state;
1278 * # RPCs enqueued for later dispatching by the policy
1282 * # RPCs started for dispatch by the policy
1284 long pi_req_started;
1286 * Is this a fallback policy?
1288 unsigned pi_fallback:1;
1294 * There is one instance of this for each policy in each NRS head of each
1295 * PTLRPC service partition.
1297 struct ptlrpc_nrs_policy {
1299 * Linkage into the NRS head's list of policies,
1300 * ptlrpc_nrs:nrs_policy_list
1302 struct list_head pol_list;
1304 * Linkage into the NRS head's list of policies with enqueued
1305 * requests ptlrpc_nrs:nrs_policy_queued
1307 struct list_head pol_list_queued;
1309 * Current state of this policy
1311 enum ptlrpc_nrs_pol_state pol_state;
1313 * Bitmask of nrs_policy_flags
1317 * # RPCs enqueued for later dispatching by the policy
1319 long pol_req_queued;
1321 * # RPCs started for dispatch by the policy
1323 long pol_req_started;
1325 * Usage Reference count taken on the policy instance
1329 * The NRS head this policy has been created at
1331 struct ptlrpc_nrs *pol_nrs;
1333 * Private policy data; varies by policy type
1337 * Policy descriptor for this policy instance.
1339 struct ptlrpc_nrs_pol_desc *pol_desc;
1345 * Resources are embedded into two types of NRS entities:
1346 * - Inside NRS policies, in the policy's private data in
1347 * ptlrpc_nrs_policy::pol_private
1348 * - In objects that act as prime-level scheduling entities in different NRS
1349 * policies; e.g. on a policy that performs round robin or similar order
1350 * scheduling across client NIDs, there would be one NRS resource per unique
1351 * client NID. On a policy which performs round robin scheduling across
1352 * backend filesystem objects, there would be one resource associated with
1353 * each of the backend filesystem objects partaking in the scheduling
1354 * performed by the policy.
1356 * NRS resources share a parent-child relationship, in which resources embedded
1357 * in policy instances are the parent entities, with all scheduling entities
1358 * a policy schedules across being the children, thus forming a simple resource
1359 * hierarchy. This hierarchy may be extended with one or more levels in the
1360 * future if the ability to have more than one primary policy is added.
1362 * Upon request initialization, references to the then active NRS policies are
1363 * taken and used to later handle the dispatching of the request with one of
1366 * \see nrs_resource_get_safe()
1367 * \see ptlrpc_nrs_req_add()
1369 struct ptlrpc_nrs_resource {
1371 * This NRS resource's parent; is NULL for resources embedded in NRS
1372 * policy instances; i.e. those are top-level ones.
1374 struct ptlrpc_nrs_resource *res_parent;
1376 * The policy associated with this resource.
1378 struct ptlrpc_nrs_policy *res_policy;
1391 * This policy is a logical wrapper around previous, non-NRS functionality.
1392 * It dispatches RPCs in the same order as they arrive from the network. This
1393 * policy is currently used as the fallback policy, and the only enabled policy
1394 * on all NRS heads of all PTLRPC service partitions.
1399 * Private data structure for the FIFO policy
1401 struct nrs_fifo_head {
1403 * Resource object for policy instance.
1405 struct ptlrpc_nrs_resource fh_res;
1407 * List of queued requests.
1409 struct list_head fh_list;
1411 * For debugging purposes.
1416 struct nrs_fifo_req {
1417 struct list_head fr_list;
1426 * CRR-N, Client Round Robin over NIDs
1431 * private data structure for CRR-N NRS
1433 struct nrs_crrn_net {
1434 struct ptlrpc_nrs_resource cn_res;
1435 cfs_binheap_t *cn_binheap;
1436 cfs_hash_t *cn_cli_hash;
1438 * Used when a new scheduling round commences, in order to synchronize
1439 * all clients with the new round number.
1443 * Determines the relevant ordering amongst request batches within a
1448 * Round Robin quantum; the maximum number of RPCs that each request
1449 * batch for each client can have in a scheduling round.
1455 * Object representing a client in CRR-N, as identified by its NID
1457 struct nrs_crrn_client {
1458 struct ptlrpc_nrs_resource cc_res;
1459 struct hlist_node cc_hnode;
1462 * The round number against which this client is currently scheduling
1467 * The sequence number used for requests scheduled by this client during
1468 * the current round number.
1473 * Round Robin quantum; the maximum number of RPCs the client is allowed
1474 * to schedule in a single batch of each round.
1478 * # of pending requests for this client, on all existing rounds
1484 * CRR-N NRS request definition
1486 struct nrs_crrn_req {
1488 * Round number for this request; shared with all other requests in the
1493 * Sequence number for this request; shared with all other requests in
1500 * CRR-N policy operations.
1504 * Read the RR quantum size of a CRR-N policy.
1506 NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
1508 * Write the RR quantum size of a CRR-N policy.
1510 NRS_CTL_CRRN_WR_QUANTUM,
1518 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1523 * Lower and upper byte offsets of a brw RPC
1525 struct nrs_orr_req_range {
1531 * RPC types supported by the ORR/TRR policies
1534 NOS_OST_READ = (1 << 0),
1535 NOS_OST_WRITE = (1 << 1),
1536 NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE),
1538 * Default value for policies.
1540 NOS_DFLT = NOS_OST_READ
1544 * As unique keys for grouping RPCs together, we use the object's OST FID for
1545 * the ORR policy, and the OST index for the TRR policy.
1547 * XXX: We waste some space for TRR policy instances by using a union, but it
1548 * allows to consolidate some of the code between ORR and TRR, and these
1549 * policies will probably eventually merge into one anyway.
1551 struct nrs_orr_key {
1553 /** object FID for ORR */
1554 struct lu_fid ok_fid;
1555 /** OST index for TRR */
1561 * The largest base string for unique hash/slab object names is
1562 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1563 * id number, so this _should_ be more than enough for the maximum number of
1564 * CPTs on any system. If it does happen that this statement is incorrect,
1565 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1566 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1567 * will hopefully not go unnoticed.
1569 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1572 * private data structure for ORR and TRR NRS
1574 struct nrs_orr_data {
1575 struct ptlrpc_nrs_resource od_res;
1576 cfs_binheap_t *od_binheap;
1577 cfs_hash_t *od_obj_hash;
1578 struct kmem_cache *od_cache;
1580 * Used when a new scheduling round commences, in order to synchronize
1581 * all object or OST batches with the new round number.
1585 * Determines the relevant ordering amongst request batches within a
1590 * RPC types that are currently supported.
1592 enum nrs_orr_supp od_supp;
1594 * Round Robin quantum; the maxium number of RPCs that each request
1595 * batch for each object or OST can have in a scheduling round.
1599 * Whether to use physical disk offsets or logical file offsets.
1603 * XXX: We need to provide a persistently allocated string to hold
1604 * unique object names for this policy, since in currently supported
1605 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1606 * to the name string provided. kstrdup() is used in the version of
1607 * kmeme_cache_create() in current Linux mainline, so we may be able to
1608 * remove this in the future.
1610 char od_objname[NRS_ORR_OBJ_NAME_MAX];
1614 * Represents a backend-fs object or OST in the ORR and TRR policies
1617 struct nrs_orr_object {
1618 struct ptlrpc_nrs_resource oo_res;
1619 struct hlist_node oo_hnode;
1621 * The round number against which requests are being scheduled for this
1626 * The sequence number used for requests scheduled for this object or
1627 * OST during the current round number.
1631 * The key of the object or OST for which this structure instance is
1634 struct nrs_orr_key oo_key;
1637 * Round Robin quantum; the maximum number of RPCs that are allowed to
1638 * be scheduled for the object or OST in a single batch of each round.
1642 * # of pending requests for this object or OST, on all existing rounds
1648 * ORR/TRR NRS request definition
1650 struct nrs_orr_req {
1652 * The offset range this request covers
1654 struct nrs_orr_req_range or_range;
1656 * Round number for this request; shared with all other requests in the
1661 * Sequence number for this request; shared with all other requests in
1666 * For debugging purposes.
1668 struct nrs_orr_key or_key;
1670 * An ORR policy instance has filled in request information while
1671 * enqueueing the request on the service partition's regular NRS head.
1673 unsigned int or_orr_set:1;
1675 * A TRR policy instance has filled in request information while
1676 * enqueueing the request on the service partition's regular NRS head.
1678 unsigned int or_trr_set:1;
1680 * Request offset ranges have been filled in with logical offset
1683 unsigned int or_logical_set:1;
1685 * Request offset ranges have been filled in with physical offset
1688 unsigned int or_physical_set:1;
1693 #include <lustre_nrs_tbf.h>
1698 * Instances of this object exist embedded within ptlrpc_request; the main
1699 * purpose of this object is to hold references to the request's resources
1700 * for the lifetime of the request, and to hold properties that policies use
1701 * use for determining the request's scheduling priority.
1703 struct ptlrpc_nrs_request {
1705 * The request's resource hierarchy.
1707 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1709 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1710 * policy that was used to enqueue the request.
1712 * \see nrs_request_enqueue()
1714 unsigned nr_res_idx;
1715 unsigned nr_initialized:1;
1716 unsigned nr_enqueued:1;
1717 unsigned nr_started:1;
1718 unsigned nr_finalized:1;
1719 cfs_binheap_node_t nr_node;
1722 * Policy-specific fields, used for determining a request's scheduling
1723 * priority, and other supporting functionality.
1727 * Fields for the FIFO policy
1729 struct nrs_fifo_req fifo;
1731 * CRR-N request defintion
1733 struct nrs_crrn_req crr;
1734 /** ORR and TRR share the same request definition */
1735 struct nrs_orr_req orr;
1737 * TBF request definition
1739 struct nrs_tbf_req tbf;
1742 * Externally-registering policies may want to use this to allocate
1743 * their own request properties.
1751 * Basic request prioritization operations structure.
1752 * The whole idea is centered around locks and RPCs that might affect locks.
1753 * When a lock is contended we try to give priority to RPCs that might lead
1754 * to fastest release of that lock.
1755 * Currently only implemented for OSTs only in a way that makes all
1756 * IO and truncate RPCs that are coming from a locked region where a lock is
1757 * contended a priority over other requests.
1759 struct ptlrpc_hpreq_ops {
1761 * Check if the lock handle of the given lock is the same as
1762 * taken from the request.
1764 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1766 * Check if the request is a high priority one.
1768 int (*hpreq_check)(struct ptlrpc_request *);
1770 * Called after the request has been handled.
1772 void (*hpreq_fini)(struct ptlrpc_request *);
1776 * Represents remote procedure call.
1778 * This is a staple structure used by everybody wanting to send a request
1781 struct ptlrpc_request {
1782 /* Request type: one of PTL_RPC_MSG_* */
1784 /** Result of request processing */
1787 * Linkage item through which this request is included into
1788 * sending/delayed lists on client and into rqbd list on server
1790 struct list_head rq_list;
1792 * Server side list of incoming unserved requests sorted by arrival
1793 * time. Traversed from time to time to notice about to expire
1794 * requests and sent back "early replies" to clients to let them
1795 * know server is alive and well, just very busy to service their
1798 struct list_head rq_timed_list;
1799 /** server-side history, used for debuging purposes. */
1800 struct list_head rq_history_list;
1801 /** server-side per-export list */
1802 struct list_head rq_exp_list;
1803 /** server-side hp handlers */
1804 struct ptlrpc_hpreq_ops *rq_ops;
1806 /** initial thread servicing this request */
1807 struct ptlrpc_thread *rq_svc_thread;
1809 /** history sequence # */
1810 __u64 rq_history_seq;
1814 /** stub for NRS request */
1815 struct ptlrpc_nrs_request rq_nrq;
1817 /** the index of service's srv_at_array into which request is linked */
1819 /** Lock to protect request flags and some other important bits, like
1823 /** client-side flags are serialized by rq_lock */
1824 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1825 rq_timedout:1, rq_resend:1, rq_restart:1,
1827 * when ->rq_replay is set, request is kept by the client even
1828 * after server commits corresponding transaction. This is
1829 * used for operations that require sequence of multiple
1830 * requests to be replayed. The only example currently is file
1831 * open/close. When last request in such a sequence is
1832 * committed, ->rq_replay is cleared on all requests in the
1836 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1837 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1839 rq_req_unlink:1, rq_reply_unlink:1,
1840 rq_memalloc:1, /* req originated from "kswapd" */
1841 /* server-side flags */
1842 rq_packed_final:1, /* packed final reply */
1843 rq_hp:1, /* high priority RPC */
1844 rq_at_linked:1, /* link into service's srv_at_array */
1845 rq_reply_truncate:1,
1847 /* whether the "rq_set" is a valid one */
1849 rq_generation_set:1,
1850 /* do not resend request on -EINPROGRESS */
1851 rq_no_retry_einprogress:1,
1852 /* allow the req to be sent if the import is in recovery
1855 /* bulk request, sent to server, but uncommitted */
1858 unsigned int rq_nr_resend;
1860 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1861 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1862 atomic_t rq_refcount;/* client-side refcount for SENT race,
1863 server-side refcounf for multiple replies */
1865 /** Portal to which this request would be sent */
1866 short rq_request_portal; /* XXX FIXME bug 249 */
1867 /** Portal where to wait for reply and where reply would be sent */
1868 short rq_reply_portal; /* XXX FIXME bug 249 */
1872 * !rq_truncate : # reply bytes actually received,
1873 * rq_truncate : required repbuf_len for resend
1875 int rq_nob_received;
1876 /** Request length */
1880 /** Request message - what client sent */
1881 struct lustre_msg *rq_reqmsg;
1882 /** Reply message - server response */
1883 struct lustre_msg *rq_repmsg;
1884 /** Transaction number */
1889 * List item to for replay list. Not yet commited requests get linked
1891 * Also see \a rq_replay comment above.
1893 struct list_head rq_replay_list;
1896 * security and encryption data
1898 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1899 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1900 struct list_head rq_ctx_chain; /**< link to waited ctx */
1902 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1903 enum lustre_sec_part rq_sp_from;
1905 /* client/server security flags */
1907 rq_ctx_init:1, /* context initiation */
1908 rq_ctx_fini:1, /* context destroy */
1909 rq_bulk_read:1, /* request bulk read */
1910 rq_bulk_write:1, /* request bulk write */
1911 /* server authentication flags */
1912 rq_auth_gss:1, /* authenticated by gss */
1913 rq_auth_remote:1, /* authed as remote user */
1914 rq_auth_usr_root:1, /* authed as root */
1915 rq_auth_usr_mdt:1, /* authed as mdt */
1916 rq_auth_usr_ost:1, /* authed as ost */
1917 /* security tfm flags */
1920 /* doesn't expect reply FIXME */
1922 rq_pill_init:1; /* pill initialized */
1924 uid_t rq_auth_uid; /* authed uid */
1925 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1927 /* (server side), pointed directly into req buffer */
1928 struct ptlrpc_user_desc *rq_user_desc;
1930 /* various buffer pointers */
1931 struct lustre_msg *rq_reqbuf; /* req wrapper */
1932 char *rq_repbuf; /* rep buffer */
1933 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1934 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1935 int rq_reqbuf_len; /* req wrapper buf len */
1936 int rq_reqdata_len; /* req wrapper msg len */
1937 int rq_repbuf_len; /* rep buffer len */
1938 int rq_repdata_len; /* rep wrapper msg len */
1939 int rq_clrbuf_len; /* only in priv mode */
1940 int rq_clrdata_len; /* only in priv mode */
1942 /** early replies go to offset 0, regular replies go after that */
1943 unsigned int rq_reply_off;
1947 /** Fields that help to see if request and reply were swabbed or not */
1948 __u32 rq_req_swab_mask;
1949 __u32 rq_rep_swab_mask;
1951 /** What was import generation when this request was sent */
1952 int rq_import_generation;
1953 enum lustre_imp_state rq_send_state;
1955 /** how many early replies (for stats) */
1958 /** client+server request */
1959 lnet_handle_md_t rq_req_md_h;
1960 struct ptlrpc_cb_id rq_req_cbid;
1961 /** optional time limit for send attempts */
1962 cfs_duration_t rq_delay_limit;
1963 /** time request was first queued */
1964 cfs_time_t rq_queued_time;
1966 /* server-side... */
1967 /** request arrival time */
1968 struct timeval rq_arrival_time;
1969 /** separated reply state */
1970 struct ptlrpc_reply_state *rq_reply_state;
1971 /** incoming request buffer */
1972 struct ptlrpc_request_buffer_desc *rq_rqbd;
1974 /** client-only incoming reply */
1975 lnet_handle_md_t rq_reply_md_h;
1976 wait_queue_head_t rq_reply_waitq;
1977 struct ptlrpc_cb_id rq_reply_cbid;
1981 /** Peer description (the other side) */
1982 lnet_process_id_t rq_peer;
1983 /** Server-side, export on which request was received */
1984 struct obd_export *rq_export;
1985 /** Client side, import where request is being sent */
1986 struct obd_import *rq_import;
1988 /** Replay callback, called after request is replayed at recovery */
1989 void (*rq_replay_cb)(struct ptlrpc_request *);
1991 * Commit callback, called when request is committed and about to be
1994 void (*rq_commit_cb)(struct ptlrpc_request *);
1995 /** Opaq data for replay and commit callbacks. */
1998 /** For bulk requests on client only: bulk descriptor */
1999 struct ptlrpc_bulk_desc *rq_bulk;
2001 /** client outgoing req */
2003 * when request/reply sent (secs), or time when request should be sent
2006 /** time for request really sent out */
2007 time_t rq_real_sent;
2009 /** when request must finish. volatile
2010 * so that servers' early reply updates to the deadline aren't
2011 * kept in per-cpu cache */
2012 volatile time_t rq_deadline;
2013 /** when req reply unlink must finish. */
2014 time_t rq_reply_deadline;
2015 /** when req bulk unlink must finish. */
2016 time_t rq_bulk_deadline;
2018 * service time estimate (secs)
2019 * If the requestsis not served by this time, it is marked as timed out.
2023 /** Multi-rpc bits */
2024 /** Per-request waitq introduced by bug 21938 for recovery waiting */
2025 wait_queue_head_t rq_set_waitq;
2026 /** Link item for request set lists */
2027 struct list_head rq_set_chain;
2028 /** Link back to the request set */
2029 struct ptlrpc_request_set *rq_set;
2030 /** Async completion handler, called when reply is received */
2031 ptlrpc_interpterer_t rq_interpret_reply;
2032 /** Resend handler, called when request is resend to update RPC data */
2033 ptlrpc_resend_cb_t rq_resend_cb;
2034 /** Async completion context */
2035 union ptlrpc_async_args rq_async_args;
2037 /** Pool if request is from preallocated list */
2038 struct ptlrpc_request_pool *rq_pool;
2040 struct lu_context rq_session;
2042 /** request format description */
2043 struct req_capsule rq_pill;
2047 * Call completion handler for rpc if any, return it's status or original
2048 * rc if there was no handler defined for this request.
2050 static inline int ptlrpc_req_interpret(const struct lu_env *env,
2051 struct ptlrpc_request *req, int rc)
2053 if (req->rq_interpret_reply != NULL) {
2054 req->rq_status = req->rq_interpret_reply(env, req,
2055 &req->rq_async_args,
2057 return req->rq_status;
2065 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
2066 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
2067 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
2068 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
2069 struct ptlrpc_nrs_pol_info *info);
2072 * Can the request be moved from the regular NRS head to the high-priority NRS
2073 * head (of the same PTLRPC service partition), if any?
2075 * For a reliable result, this should be checked under svcpt->scp_req lock.
2077 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
2079 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
2082 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
2083 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
2084 * to make sure it has not been scheduled yet (analogous to previous
2085 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
2087 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
2092 * Returns 1 if request buffer at offset \a index was already swabbed
2094 static inline int lustre_req_swabbed(struct ptlrpc_request *req, size_t index)
2096 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2097 return req->rq_req_swab_mask & (1 << index);
2101 * Returns 1 if request reply buffer at offset \a index was already swabbed
2103 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, size_t index)
2105 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2106 return req->rq_rep_swab_mask & (1 << index);
2110 * Returns 1 if request needs to be swabbed into local cpu byteorder
2112 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
2114 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2118 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
2120 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
2122 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
2126 * Mark request buffer at offset \a index that it was already swabbed
2128 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req,
2131 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
2132 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
2133 req->rq_req_swab_mask |= 1 << index;
2137 * Mark request reply buffer at offset \a index that it was already swabbed
2139 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req,
2142 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
2143 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
2144 req->rq_rep_swab_mask |= 1 << index;
2148 * Convert numerical request phase value \a phase into text string description
2150 static inline const char *
2151 ptlrpc_phase2str(enum rq_phase phase)
2160 case RQ_PHASE_INTERPRET:
2162 case RQ_PHASE_COMPLETE:
2164 case RQ_PHASE_UNREGISTERING:
2165 return "Unregistering";
2172 * Convert numerical request phase of the request \a req into text stringi
2175 static inline const char *
2176 ptlrpc_rqphase2str(struct ptlrpc_request *req)
2178 return ptlrpc_phase2str(req->rq_phase);
2182 * Debugging functions and helpers to print request structure into debug log
2185 /* Spare the preprocessor, spoil the bugs. */
2186 #define FLAG(field, str) (field ? str : "")
2188 /** Convert bit flags into a string */
2189 #define DEBUG_REQ_FLAGS(req) \
2190 ptlrpc_rqphase2str(req), \
2191 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
2192 FLAG(req->rq_err, "E"), \
2193 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
2194 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
2195 FLAG(req->rq_no_resend, "N"), \
2196 FLAG(req->rq_waiting, "W"), \
2197 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
2198 FLAG(req->rq_committed, "M")
2200 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
2202 void _debug_req(struct ptlrpc_request *req,
2203 struct libcfs_debug_msg_data *data, const char *fmt, ...)
2204 __attribute__ ((format (printf, 3, 4)));
2207 * Helper that decides if we need to print request accordig to current debug
2210 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
2212 CFS_CHECK_STACK(msgdata, mask, cdls); \
2214 if (((mask) & D_CANTMASK) != 0 || \
2215 ((libcfs_debug & (mask)) != 0 && \
2216 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
2217 _debug_req((req), msgdata, fmt, ##a); \
2221 * This is the debug print function you need to use to print request sturucture
2222 * content into lustre debug log.
2223 * for most callers (level is a constant) this is resolved at compile time */
2224 #define DEBUG_REQ(level, req, fmt, args...) \
2226 if ((level) & (D_ERROR | D_WARNING)) { \
2227 static cfs_debug_limit_state_t cdls; \
2228 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
2229 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
2231 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
2232 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
2238 * Structure that defines a single page of a bulk transfer
2240 struct ptlrpc_bulk_page {
2241 /** Linkage to list of pages in a bulk */
2242 struct list_head bp_link;
2244 * Number of bytes in a page to transfer starting from \a bp_pageoffset
2247 /** offset within a page */
2249 /** The page itself */
2250 struct page *bp_page;
2253 #define BULK_GET_SOURCE 0
2254 #define BULK_PUT_SINK 1
2255 #define BULK_GET_SINK 2
2256 #define BULK_PUT_SOURCE 3
2259 * Definition of bulk descriptor.
2260 * Bulks are special "Two phase" RPCs where initial request message
2261 * is sent first and it is followed bt a transfer (o receiving) of a large
2262 * amount of data to be settled into pages referenced from the bulk descriptors.
2263 * Bulks transfers (the actual data following the small requests) are done
2264 * on separate LNet portals.
2265 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2266 * Another user is readpage for MDT.
2268 struct ptlrpc_bulk_desc {
2269 /** completed with failure */
2270 unsigned long bd_failure:1;
2271 /** {put,get}{source,sink} */
2272 unsigned long bd_type:2;
2274 unsigned long bd_registered:1;
2275 /** For serialization with callback */
2277 /** Import generation when request for this bulk was sent */
2278 int bd_import_generation;
2279 /** LNet portal for this bulk */
2281 /** Server side - export this bulk created for */
2282 struct obd_export *bd_export;
2283 /** Client side - import this bulk was sent on */
2284 struct obd_import *bd_import;
2285 /** Back pointer to the request */
2286 struct ptlrpc_request *bd_req;
2287 wait_queue_head_t bd_waitq; /* server side only WQ */
2288 int bd_iov_count; /* # entries in bd_iov */
2289 int bd_max_iov; /* allocated size of bd_iov */
2290 int bd_nob; /* # bytes covered */
2291 int bd_nob_transferred; /* # bytes GOT/PUT */
2295 struct ptlrpc_cb_id bd_cbid; /* network callback info */
2296 lnet_nid_t bd_sender; /* stash event::sender */
2297 int bd_md_count; /* # valid entries in bd_mds */
2298 int bd_md_max_brw; /* max entries in bd_mds */
2299 /** array of associated MDs */
2300 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
2303 * encrypt iov, size is either 0 or bd_iov_count.
2305 lnet_kiov_t *bd_enc_iov;
2307 lnet_kiov_t bd_iov[0];
2311 SVC_STOPPED = 1 << 0,
2312 SVC_STOPPING = 1 << 1,
2313 SVC_STARTING = 1 << 2,
2314 SVC_RUNNING = 1 << 3,
2316 SVC_SIGNAL = 1 << 5,
2319 #define PTLRPC_THR_NAME_LEN 32
2321 * Definition of server service thread structure
2323 struct ptlrpc_thread {
2325 * List of active threads in svc->srv_threads
2327 struct list_head t_link;
2329 * thread-private data (preallocated memory)
2334 * service thread index, from ptlrpc_start_threads
2338 * service thread pid
2342 * put watchdog in the structure per thread b=14840
2344 struct lc_watchdog *t_watchdog;
2346 * the svc this thread belonged to b=18582
2348 struct ptlrpc_service_part *t_svcpt;
2349 wait_queue_head_t t_ctl_waitq;
2350 struct lu_env *t_env;
2351 char t_name[PTLRPC_THR_NAME_LEN];
2354 static inline int thread_is_init(struct ptlrpc_thread *thread)
2356 return thread->t_flags == 0;
2359 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
2361 return !!(thread->t_flags & SVC_STOPPED);
2364 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
2366 return !!(thread->t_flags & SVC_STOPPING);
2369 static inline int thread_is_starting(struct ptlrpc_thread *thread)
2371 return !!(thread->t_flags & SVC_STARTING);
2374 static inline int thread_is_running(struct ptlrpc_thread *thread)
2376 return !!(thread->t_flags & SVC_RUNNING);
2379 static inline int thread_is_event(struct ptlrpc_thread *thread)
2381 return !!(thread->t_flags & SVC_EVENT);
2384 static inline int thread_is_signal(struct ptlrpc_thread *thread)
2386 return !!(thread->t_flags & SVC_SIGNAL);
2389 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
2391 thread->t_flags &= ~flags;
2394 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
2396 thread->t_flags = flags;
2399 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
2401 thread->t_flags |= flags;
2404 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
2407 if (thread->t_flags & flags) {
2408 thread->t_flags &= ~flags;
2415 * Request buffer descriptor structure.
2416 * This is a structure that contains one posted request buffer for service.
2417 * Once data land into a buffer, event callback creates actual request and
2418 * notifies wakes one of the service threads to process new incoming request.
2419 * More than one request can fit into the buffer.
2421 struct ptlrpc_request_buffer_desc {
2422 /** Link item for rqbds on a service */
2423 struct list_head rqbd_list;
2424 /** History of requests for this buffer */
2425 struct list_head rqbd_reqs;
2426 /** Back pointer to service for which this buffer is registered */
2427 struct ptlrpc_service_part *rqbd_svcpt;
2428 /** LNet descriptor */
2429 lnet_handle_md_t rqbd_md_h;
2431 /** The buffer itself */
2433 struct ptlrpc_cb_id rqbd_cbid;
2435 * This "embedded" request structure is only used for the
2436 * last request to fit into the buffer
2438 struct ptlrpc_request rqbd_req;
2441 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2443 struct ptlrpc_service_ops {
2445 * if non-NULL called during thread creation (ptlrpc_start_thread())
2446 * to initialize service specific per-thread state.
2448 int (*so_thr_init)(struct ptlrpc_thread *thr);
2450 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2451 * destruct state created by ->srv_init().
2453 void (*so_thr_done)(struct ptlrpc_thread *thr);
2455 * Handler function for incoming requests for this service
2457 int (*so_req_handler)(struct ptlrpc_request *req);
2459 * function to determine priority of the request, it's called
2460 * on every new request
2462 int (*so_hpreq_handler)(struct ptlrpc_request *);
2464 * service-specific print fn
2466 void (*so_req_printer)(void *, struct ptlrpc_request *);
2469 #ifndef __cfs_cacheline_aligned
2470 /* NB: put it here for reducing patche dependence */
2471 # define __cfs_cacheline_aligned
2475 * How many high priority requests to serve before serving one normal
2478 #define PTLRPC_SVC_HP_RATIO 10
2481 * Definition of PortalRPC service.
2482 * The service is listening on a particular portal (like tcp port)
2483 * and perform actions for a specific server like IO service for OST
2484 * or general metadata service for MDS.
2486 struct ptlrpc_service {
2487 /** serialize /proc operations */
2488 spinlock_t srv_lock;
2489 /** most often accessed fields */
2490 /** chain thru all services */
2491 struct list_head srv_list;
2492 /** service operations table */
2493 struct ptlrpc_service_ops srv_ops;
2494 /** only statically allocated strings here; we don't clean them */
2496 /** only statically allocated strings here; we don't clean them */
2497 char *srv_thread_name;
2498 /** service thread list */
2499 struct list_head srv_threads;
2500 /** threads # should be created for each partition on initializing */
2501 int srv_nthrs_cpt_init;
2502 /** limit of threads number for each partition */
2503 int srv_nthrs_cpt_limit;
2504 /** Root of /proc dir tree for this service */
2505 struct proc_dir_entry *srv_procroot;
2506 /** Pointer to statistic data for this service */
2507 struct lprocfs_stats *srv_stats;
2508 /** # hp per lp reqs to handle */
2509 int srv_hpreq_ratio;
2510 /** biggest request to receive */
2511 int srv_max_req_size;
2512 /** biggest reply to send */
2513 int srv_max_reply_size;
2514 /** size of individual buffers */
2516 /** # buffers to allocate in 1 group */
2517 int srv_nbuf_per_group;
2518 /** Local portal on which to receive requests */
2519 __u32 srv_req_portal;
2520 /** Portal on the client to send replies to */
2521 __u32 srv_rep_portal;
2523 * Tags for lu_context associated with this thread, see struct
2527 /** soft watchdog timeout multiplier */
2528 int srv_watchdog_factor;
2529 /** under unregister_service */
2530 unsigned srv_is_stopping:1;
2532 /** max # request buffers in history per partition */
2533 int srv_hist_nrqbds_cpt_max;
2534 /** number of CPTs this service bound on */
2536 /** CPTs array this service bound on */
2538 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2540 /** CPT table this service is running over */
2541 struct cfs_cpt_table *srv_cptable;
2543 * partition data for ptlrpc service
2545 struct ptlrpc_service_part *srv_parts[0];
2549 * Definition of PortalRPC service partition data.
2550 * Although a service only has one instance of it right now, but we
2551 * will have multiple instances very soon (instance per CPT).
2553 * it has four locks:
2555 * serialize operations on rqbd and requests waiting for preprocess
2557 * serialize operations active requests sent to this portal
2559 * serialize adaptive timeout stuff
2561 * serialize operations on RS list (reply states)
2563 * We don't have any use-case to take two or more locks at the same time
2564 * for now, so there is no lock order issue.
2566 struct ptlrpc_service_part {
2567 /** back reference to owner */
2568 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2569 /* CPT id, reserved */
2571 /** always increasing number */
2573 /** # of starting threads */
2574 int scp_nthrs_starting;
2575 /** # of stopping threads, reserved for shrinking threads */
2576 int scp_nthrs_stopping;
2577 /** # running threads */
2578 int scp_nthrs_running;
2579 /** service threads list */
2580 struct list_head scp_threads;
2583 * serialize the following fields, used for protecting
2584 * rqbd list and incoming requests waiting for preprocess,
2585 * threads starting & stopping are also protected by this lock.
2587 spinlock_t scp_lock __cfs_cacheline_aligned;
2588 /** total # req buffer descs allocated */
2589 int scp_nrqbds_total;
2590 /** # posted request buffers for receiving */
2591 int scp_nrqbds_posted;
2592 /** in progress of allocating rqbd */
2593 int scp_rqbd_allocating;
2594 /** # incoming reqs */
2595 int scp_nreqs_incoming;
2596 /** request buffers to be reposted */
2597 struct list_head scp_rqbd_idle;
2598 /** req buffers receiving */
2599 struct list_head scp_rqbd_posted;
2600 /** incoming reqs */
2601 struct list_head scp_req_incoming;
2602 /** timeout before re-posting reqs, in tick */
2603 cfs_duration_t scp_rqbd_timeout;
2605 * all threads sleep on this. This wait-queue is signalled when new
2606 * incoming request arrives and when difficult reply has to be handled.
2608 wait_queue_head_t scp_waitq;
2610 /** request history */
2611 struct list_head scp_hist_reqs;
2612 /** request buffer history */
2613 struct list_head scp_hist_rqbds;
2614 /** # request buffers in history */
2615 int scp_hist_nrqbds;
2616 /** sequence number for request */
2618 /** highest seq culled from history */
2619 __u64 scp_hist_seq_culled;
2622 * serialize the following fields, used for processing requests
2623 * sent to this portal
2625 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2626 /** # reqs in either of the NRS heads below */
2627 /** # reqs being served */
2628 int scp_nreqs_active;
2629 /** # HPreqs being served */
2630 int scp_nhreqs_active;
2631 /** # hp requests handled */
2634 /** NRS head for regular requests */
2635 struct ptlrpc_nrs scp_nrs_reg;
2636 /** NRS head for HP requests; this is only valid for services that can
2637 * handle HP requests */
2638 struct ptlrpc_nrs *scp_nrs_hp;
2643 * serialize the following fields, used for changes on
2646 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2647 /** estimated rpc service time */
2648 struct adaptive_timeout scp_at_estimate;
2649 /** reqs waiting for replies */
2650 struct ptlrpc_at_array scp_at_array;
2651 /** early reply timer */
2652 struct timer_list scp_at_timer;
2654 cfs_time_t scp_at_checktime;
2655 /** check early replies */
2656 unsigned scp_at_check;
2660 * serialize the following fields, used for processing
2661 * replies for this portal
2663 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2664 /** all the active replies */
2665 struct list_head scp_rep_active;
2666 /** List of free reply_states */
2667 struct list_head scp_rep_idle;
2668 /** waitq to run, when adding stuff to srv_free_rs_list */
2669 wait_queue_head_t scp_rep_waitq;
2670 /** # 'difficult' replies */
2671 atomic_t scp_nreps_difficult;
2674 #define ptlrpc_service_for_each_part(part, i, svc) \
2676 i < (svc)->srv_ncpts && \
2677 (svc)->srv_parts != NULL && \
2678 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2681 * Declaration of ptlrpcd control structure
2683 struct ptlrpcd_ctl {
2685 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2687 unsigned long pc_flags;
2689 * Thread lock protecting structure fields.
2695 struct completion pc_starting;
2699 struct completion pc_finishing;
2701 * Thread requests set.
2703 struct ptlrpc_request_set *pc_set;
2705 * Thread name used in kthread_run()
2709 * Environment for request interpreters to run in.
2711 struct lu_env pc_env;
2713 * Index of ptlrpcd thread in the array.
2717 * Number of the ptlrpcd's partners.
2721 * Pointer to the array of partners' ptlrpcd_ctl structure.
2723 struct ptlrpcd_ctl **pc_partners;
2725 * Record the partner index to be processed next.
2730 /* Bits for pc_flags */
2731 enum ptlrpcd_ctl_flags {
2733 * Ptlrpc thread start flag.
2735 LIOD_START = 1 << 0,
2737 * Ptlrpc thread stop flag.
2741 * Ptlrpc thread force flag (only stop force so far).
2742 * This will cause aborting any inflight rpcs handled
2743 * by thread if LIOD_STOP is specified.
2745 LIOD_FORCE = 1 << 2,
2747 * This is a recovery ptlrpc thread.
2749 LIOD_RECOVERY = 1 << 3,
2751 * The ptlrpcd is bound to some CPU core.
2760 * Service compatibility function; the policy is compatible with all services.
2762 * \param[in] svc The service the policy is attempting to register with.
2763 * \param[in] desc The policy descriptor
2765 * \retval true The policy is compatible with the service
2767 * \see ptlrpc_nrs_pol_desc::pd_compat()
2769 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2770 const struct ptlrpc_nrs_pol_desc *desc)
2776 * Service compatibility function; the policy is compatible with only a specific
2777 * service which is identified by its human-readable name at
2778 * ptlrpc_service::srv_name.
2780 * \param[in] svc The service the policy is attempting to register with.
2781 * \param[in] desc The policy descriptor
2783 * \retval false The policy is not compatible with the service
2784 * \retval true The policy is compatible with the service
2786 * \see ptlrpc_nrs_pol_desc::pd_compat()
2788 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2789 const struct ptlrpc_nrs_pol_desc *desc)
2791 LASSERT(desc->pd_compat_svc_name != NULL);
2792 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2797 /* ptlrpc/events.c */
2798 extern lnet_handle_eq_t ptlrpc_eq_h;
2799 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2800 lnet_process_id_t *peer, lnet_nid_t *self);
2802 * These callbacks are invoked by LNet when something happened to
2806 extern void request_out_callback(lnet_event_t *ev);
2807 extern void reply_in_callback(lnet_event_t *ev);
2808 extern void client_bulk_callback(lnet_event_t *ev);
2809 extern void request_in_callback(lnet_event_t *ev);
2810 extern void reply_out_callback(lnet_event_t *ev);
2811 #ifdef HAVE_SERVER_SUPPORT
2812 extern void server_bulk_callback(lnet_event_t *ev);
2816 /* ptlrpc/connection.c */
2817 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2819 struct obd_uuid *uuid);
2820 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2821 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2822 int ptlrpc_connection_init(void);
2823 void ptlrpc_connection_fini(void);
2824 extern lnet_pid_t ptl_get_pid(void);
2826 /* ptlrpc/niobuf.c */
2828 * Actual interfacing with LNet to put/get/register/unregister stuff
2831 #ifdef HAVE_SERVER_SUPPORT
2832 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2833 unsigned npages, unsigned max_brw,
2834 unsigned type, unsigned portal);
2835 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2836 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2838 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2842 LASSERT(desc != NULL);
2844 spin_lock(&desc->bd_lock);
2845 rc = desc->bd_md_count;
2846 spin_unlock(&desc->bd_lock);
2851 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2852 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2854 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2856 struct ptlrpc_bulk_desc *desc;
2859 LASSERT(req != NULL);
2860 desc = req->rq_bulk;
2862 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2863 req->rq_bulk_deadline > cfs_time_current_sec())
2869 spin_lock(&desc->bd_lock);
2870 rc = desc->bd_md_count;
2871 spin_unlock(&desc->bd_lock);
2875 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2876 #define PTLRPC_REPLY_EARLY 0x02
2877 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2878 int ptlrpc_reply(struct ptlrpc_request *req);
2879 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2880 int ptlrpc_error(struct ptlrpc_request *req);
2881 void ptlrpc_resend_req(struct ptlrpc_request *request);
2882 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2883 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2884 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2887 /* ptlrpc/client.c */
2889 * Client-side portals API. Everything to send requests, receive replies,
2890 * request queues, request management, etc.
2893 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2895 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2896 struct ptlrpc_client *);
2897 void ptlrpc_cleanup_client(struct obd_import *imp);
2898 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2900 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2901 int ptlrpc_replay_req(struct ptlrpc_request *req);
2902 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2903 void ptlrpc_restart_req(struct ptlrpc_request *req);
2904 void ptlrpc_abort_inflight(struct obd_import *imp);
2905 void ptlrpc_cleanup_imp(struct obd_import *imp);
2906 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2908 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2909 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2911 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2912 set_interpreter_func fn, void *data);
2913 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2914 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2915 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2916 int ptlrpc_expired_set(void *data);
2917 void ptlrpc_interrupted_set(void *data);
2918 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2919 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2920 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2921 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2922 struct ptlrpc_request *req);
2924 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2925 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2927 struct ptlrpc_request_pool *
2928 ptlrpc_init_rq_pool(int, int,
2929 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2931 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2932 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2933 const struct req_format *format);
2934 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2935 struct ptlrpc_request_pool *,
2936 const struct req_format *format);
2937 void ptlrpc_request_free(struct ptlrpc_request *request);
2938 int ptlrpc_request_pack(struct ptlrpc_request *request,
2939 __u32 version, int opcode);
2940 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2941 const struct req_format *format,
2942 __u32 version, int opcode);
2943 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2944 __u32 version, int opcode, char **bufs,
2945 struct ptlrpc_cli_ctx *ctx);
2946 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2947 int opcode, int count, __u32 *lengths,
2949 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2950 __u32 version, int opcode,
2951 int count, __u32 *lengths, char **bufs,
2952 struct ptlrpc_request_pool *pool);
2953 void ptlrpc_req_finished(struct ptlrpc_request *request);
2954 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2955 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2956 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2957 unsigned npages, unsigned max_brw,
2958 unsigned type, unsigned portal);
2959 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2960 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2962 __ptlrpc_free_bulk(bulk, 1);
2964 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2966 __ptlrpc_free_bulk(bulk, 0);
2968 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2969 struct page *page, int pageoffset, int len, int);
2970 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2971 struct page *page, int pageoffset,
2974 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2977 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2978 struct page *page, int pageoffset,
2981 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2984 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2985 struct obd_import *imp);
2986 __u64 ptlrpc_next_xid(void);
2987 __u64 ptlrpc_sample_next_xid(void);
2988 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2990 /* Set of routines to run a function in ptlrpcd context */
2991 void *ptlrpcd_alloc_work(struct obd_import *imp,
2992 int (*cb)(const struct lu_env *, void *), void *data);
2993 void ptlrpcd_destroy_work(void *handler);
2994 int ptlrpcd_queue_work(void *handler);
2997 struct ptlrpc_service_buf_conf {
2998 /* nbufs is buffers # to allocate when growing the pool */
2999 unsigned int bc_nbufs;
3000 /* buffer size to post */
3001 unsigned int bc_buf_size;
3002 /* portal to listed for requests on */
3003 unsigned int bc_req_portal;
3004 /* portal of where to send replies to */
3005 unsigned int bc_rep_portal;
3006 /* maximum request size to be accepted for this service */
3007 unsigned int bc_req_max_size;
3008 /* maximum reply size this service can ever send */
3009 unsigned int bc_rep_max_size;
3012 struct ptlrpc_service_thr_conf {
3013 /* threadname should be 8 characters or less - 6 will be added on */
3015 /* threads increasing factor for each CPU */
3016 unsigned int tc_thr_factor;
3017 /* service threads # to start on each partition while initializing */
3018 unsigned int tc_nthrs_init;
3020 * low water of threads # upper-limit on each partition while running,
3021 * service availability may be impacted if threads number is lower
3022 * than this value. It can be ZERO if the service doesn't require
3023 * CPU affinity or there is only one partition.
3025 unsigned int tc_nthrs_base;
3026 /* "soft" limit for total threads number */
3027 unsigned int tc_nthrs_max;
3028 /* user specified threads number, it will be validated due to
3029 * other members of this structure. */
3030 unsigned int tc_nthrs_user;
3031 /* set NUMA node affinity for service threads */
3032 unsigned int tc_cpu_affinity;
3033 /* Tags for lu_context associated with service thread */
3037 struct ptlrpc_service_cpt_conf {
3038 struct cfs_cpt_table *cc_cptable;
3039 /* string pattern to describe CPTs for a service */
3043 struct ptlrpc_service_conf {
3046 /* soft watchdog timeout multiplifier to print stuck service traces */
3047 unsigned int psc_watchdog_factor;
3048 /* buffer information */
3049 struct ptlrpc_service_buf_conf psc_buf;
3050 /* thread information */
3051 struct ptlrpc_service_thr_conf psc_thr;
3052 /* CPU partition information */
3053 struct ptlrpc_service_cpt_conf psc_cpt;
3054 /* function table */
3055 struct ptlrpc_service_ops psc_ops;
3058 /* ptlrpc/service.c */
3060 * Server-side services API. Register/unregister service, request state
3061 * management, service thread management
3065 void ptlrpc_save_lock(struct ptlrpc_request *req,
3066 struct lustre_handle *lock, int mode, int no_ack);
3067 void ptlrpc_commit_replies(struct obd_export *exp);
3068 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
3069 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
3070 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
3071 struct ptlrpc_service *ptlrpc_register_service(
3072 struct ptlrpc_service_conf *conf,
3073 struct proc_dir_entry *proc_entry);
3074 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
3076 int ptlrpc_start_threads(struct ptlrpc_service *svc);
3077 int ptlrpc_unregister_service(struct ptlrpc_service *service);
3078 int liblustre_check_services(void *arg);
3079 void ptlrpc_daemonize(char *name);
3080 int ptlrpc_service_health_check(struct ptlrpc_service *);
3081 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
3082 void ptlrpc_request_change_export(struct ptlrpc_request *req,
3083 struct obd_export *export);
3084 void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay);
3086 int ptlrpc_hr_init(void);
3087 void ptlrpc_hr_fini(void);
3091 /* ptlrpc/import.c */
3096 int ptlrpc_connect_import(struct obd_import *imp);
3097 int ptlrpc_init_import(struct obd_import *imp);
3098 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
3099 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
3100 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
3103 /* ptlrpc/pack_generic.c */
3104 int ptlrpc_reconnect_import(struct obd_import *imp);
3108 * ptlrpc msg buffer and swab interface
3112 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
3114 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
3116 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
3117 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
3119 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
3120 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
3122 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
3123 __u32 *lens, char **bufs);
3124 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
3126 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
3127 __u32 *lens, char **bufs, int flags);
3128 #define LPRFL_EARLY_REPLY 1
3129 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
3130 char **bufs, int flags);
3131 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
3132 unsigned int newlen, int move_data);
3133 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
3134 int __lustre_unpack_msg(struct lustre_msg *m, int len);
3135 int lustre_msg_hdr_size(__u32 magic, int count);
3136 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
3137 int lustre_msg_size_v2(int count, __u32 *lengths);
3138 int lustre_packed_msg_size(struct lustre_msg *msg);
3139 int lustre_msg_early_size(void);
3140 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
3141 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
3142 int lustre_msg_buflen(struct lustre_msg *m, int n);
3143 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
3144 int lustre_msg_bufcount(struct lustre_msg *m);
3145 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
3146 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
3147 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
3148 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
3149 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
3150 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
3151 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
3152 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
3153 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
3154 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
3155 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
3156 __u32 lustre_msg_get_type(struct lustre_msg *msg);
3157 __u32 lustre_msg_get_version(struct lustre_msg *msg);
3158 void lustre_msg_add_version(struct lustre_msg *msg, int version);
3159 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
3160 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
3161 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
3162 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
3163 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
3164 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
3165 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
3166 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
3167 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
3168 int lustre_msg_get_status(struct lustre_msg *msg);
3169 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
3170 int lustre_msg_is_v1(struct lustre_msg *msg);
3171 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
3172 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
3173 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
3174 char *lustre_msg_get_jobid(struct lustre_msg *msg);
3175 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
3176 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 53, 0)
3177 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
3179 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
3181 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
3182 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
3183 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
3184 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
3185 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
3186 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
3187 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
3188 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
3189 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
3190 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
3191 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
3192 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
3193 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
3194 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
3195 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
3198 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
3199 unsigned int newlen, int move_data)
3201 LASSERT(req->rq_reply_state);
3202 LASSERT(req->rq_repmsg);
3203 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
3207 #ifdef LUSTRE_TRANSLATE_ERRNOS
3209 static inline int ptlrpc_status_hton(int h)
3212 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
3213 * ELDLM_LOCK_ABORTED, etc.
3216 return -lustre_errno_hton(-h);
3221 static inline int ptlrpc_status_ntoh(int n)
3224 * See the comment in ptlrpc_status_hton().
3227 return -lustre_errno_ntoh(-n);
3234 #define ptlrpc_status_hton(h) (h)
3235 #define ptlrpc_status_ntoh(n) (n)
3240 /** Change request phase of \a req to \a new_phase */
3242 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
3244 if (req->rq_phase == new_phase)
3247 if (new_phase == RQ_PHASE_UNREGISTERING) {
3248 req->rq_next_phase = req->rq_phase;
3250 atomic_inc(&req->rq_import->imp_unregistering);
3253 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
3255 atomic_dec(&req->rq_import->imp_unregistering);
3258 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
3259 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
3261 req->rq_phase = new_phase;
3265 * Returns true if request \a req got early reply and hard deadline is not met
3268 ptlrpc_client_early(struct ptlrpc_request *req)
3270 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3271 req->rq_reply_deadline > cfs_time_current_sec())
3273 return req->rq_early;
3277 * Returns true if we got real reply from server for this request
3280 ptlrpc_client_replied(struct ptlrpc_request *req)
3282 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3283 req->rq_reply_deadline > cfs_time_current_sec())
3285 return req->rq_replied;
3288 /** Returns true if request \a req is in process of receiving server reply */
3290 ptlrpc_client_recv(struct ptlrpc_request *req)
3292 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3293 req->rq_reply_deadline > cfs_time_current_sec())
3295 return req->rq_receiving_reply;
3299 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
3303 spin_lock(&req->rq_lock);
3304 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
3305 req->rq_reply_deadline > cfs_time_current_sec()) {
3306 spin_unlock(&req->rq_lock);
3309 rc = req->rq_receiving_reply ;
3310 rc = rc || req->rq_req_unlink || req->rq_reply_unlink;
3311 spin_unlock(&req->rq_lock);
3316 ptlrpc_client_wake_req(struct ptlrpc_request *req)
3318 if (req->rq_set == NULL)
3319 wake_up(&req->rq_reply_waitq);
3321 wake_up(&req->rq_set->set_waitq);
3325 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
3327 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3328 atomic_inc(&rs->rs_refcount);
3332 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
3334 LASSERT(atomic_read(&rs->rs_refcount) > 0);
3335 if (atomic_dec_and_test(&rs->rs_refcount))
3336 lustre_free_reply_state(rs);
3339 /* Should only be called once per req */
3340 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
3342 if (req->rq_reply_state == NULL)
3343 return; /* shouldn't occur */
3344 ptlrpc_rs_decref(req->rq_reply_state);
3345 req->rq_reply_state = NULL;
3346 req->rq_repmsg = NULL;
3349 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
3351 return lustre_msg_get_magic(req->rq_reqmsg);
3354 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
3356 switch (req->rq_reqmsg->lm_magic) {
3357 case LUSTRE_MSG_MAGIC_V2:
3358 return req->rq_reqmsg->lm_repsize;
3360 LASSERTF(0, "incorrect message magic: %08x\n",
3361 req->rq_reqmsg->lm_magic);
3366 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
3368 if (req->rq_delay_limit != 0 &&
3369 cfs_time_before(cfs_time_add(req->rq_queued_time,
3370 cfs_time_seconds(req->rq_delay_limit)),
3371 cfs_time_current())) {
3377 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
3379 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
3380 spin_lock(&req->rq_lock);
3381 req->rq_no_resend = 1;
3382 spin_unlock(&req->rq_lock);
3384 return req->rq_no_resend;
3388 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
3390 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
3392 return svcpt->scp_service->srv_watchdog_factor *
3393 max_t(int, at, obd_timeout);
3396 static inline struct ptlrpc_service *
3397 ptlrpc_req2svc(struct ptlrpc_request *req)
3399 LASSERT(req->rq_rqbd != NULL);
3400 return req->rq_rqbd->rqbd_svcpt->scp_service;
3403 /* ldlm/ldlm_lib.c */
3405 * Target client logic
3408 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
3409 int client_obd_cleanup(struct obd_device *obddev);
3410 int client_connect_import(const struct lu_env *env,
3411 struct obd_export **exp, struct obd_device *obd,
3412 struct obd_uuid *cluuid, struct obd_connect_data *,
3414 int client_disconnect_export(struct obd_export *exp);
3415 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
3417 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
3418 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
3419 struct obd_uuid *uuid);
3420 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
3421 void client_destroy_import(struct obd_import *imp);
3424 #ifdef HAVE_SERVER_SUPPORT
3425 int server_disconnect_export(struct obd_export *exp);
3428 /* ptlrpc/pinger.c */
3430 * Pinger API (client side only)
3433 enum timeout_event {
3436 struct timeout_item;
3437 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
3438 int ptlrpc_pinger_add_import(struct obd_import *imp);
3439 int ptlrpc_pinger_del_import(struct obd_import *imp);
3440 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
3441 timeout_cb_t cb, void *data,
3442 struct list_head *obd_list);
3443 int ptlrpc_del_timeout_client(struct list_head *obd_list,
3444 enum timeout_event event);
3445 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3446 int ptlrpc_obd_ping(struct obd_device *obd);
3447 void ping_evictor_start(void);
3448 void ping_evictor_stop(void);
3449 void ptlrpc_pinger_ir_up(void);
3450 void ptlrpc_pinger_ir_down(void);
3452 int ptlrpc_pinger_suppress_pings(void);
3454 /* ptlrpc daemon bind policy */
3456 /* all ptlrpcd threads are free mode */
3457 PDB_POLICY_NONE = 1,
3458 /* all ptlrpcd threads are bound mode */
3459 PDB_POLICY_FULL = 2,
3460 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3461 PDB_POLICY_PAIR = 3,
3462 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3463 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3464 * If kernel supports NUMA, pthrpcd threads are binded and
3465 * grouped by NUMA node */
3466 PDB_POLICY_NEIGHBOR = 4,
3469 /* ptlrpc daemon load policy
3470 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3471 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3472 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3473 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3474 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3476 /* on the same CPU core as the caller */
3477 PDL_POLICY_SAME = 1,
3478 /* within the same CPU partition, but not the same core as the caller */
3479 PDL_POLICY_LOCAL = 2,
3480 /* round-robin on all CPU cores, but not the same core as the caller */
3481 PDL_POLICY_ROUND = 3,
3482 /* the specified CPU core is preferred, but not enforced */
3483 PDL_POLICY_PREFERRED = 4,
3486 /* ptlrpc/ptlrpcd.c */
3487 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3488 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
3489 void ptlrpcd_wake(struct ptlrpc_request *req);
3490 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3491 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3492 int ptlrpcd_addref(void);
3493 void ptlrpcd_decref(void);
3495 /* ptlrpc/lproc_ptlrpc.c */
3497 * procfs output related functions
3500 const char* ll_opcode2str(__u32 opcode);
3502 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3503 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3504 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3506 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3507 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3508 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3512 /* ptlrpc/llog_server.c */
3513 int llog_origin_handle_open(struct ptlrpc_request *req);
3514 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3515 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3516 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3517 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3518 int llog_origin_handle_close(struct ptlrpc_request *req);
3520 /* ptlrpc/llog_client.c */
3521 extern struct llog_operations llog_client_ops;