4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #if defined(__linux__)
59 #include <linux/lustre_net.h>
60 #elif defined(__APPLE__)
61 #include <darwin/lustre_net.h>
62 #elif defined(__WINNT__)
63 #include <winnt/lustre_net.h>
65 #error Unsupported operating system.
68 #include <libcfs/libcfs.h>
70 #include <lnet/lnet.h>
71 #include <lustre/lustre_idl.h>
72 #include <lustre_ha.h>
73 #include <lustre_sec.h>
74 #include <lustre_import.h>
75 #include <lprocfs_status.h>
76 #include <lu_object.h>
77 #include <lustre_req_layout.h>
79 #include <obd_support.h>
80 #include <lustre_ver.h>
82 /* MD flags we _always_ use */
83 #define PTLRPC_MD_OPTIONS 0
86 * Max # of bulk operations in one request.
87 * In order for the client and server to properly negotiate the maximum
88 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
89 * value. The client is free to limit the actual RPC size for any bulk
90 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
91 #define PTLRPC_BULK_OPS_BITS 2
92 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
94 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
95 * should not be used on the server at all. Otherwise, it imposes a
96 * protocol limitation on the maximum RPC size that can be used by any
97 * RPC sent to that server in the future. Instead, the server should
98 * use the negotiated per-client ocd_brw_size to determine the bulk
100 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
103 * Define maxima for bulk I/O.
105 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
106 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
107 * currently supported maximum between peers at connect via ocd_brw_size.
109 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
110 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
111 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
113 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
114 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
115 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
116 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
117 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> CFS_PAGE_SHIFT)
118 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
120 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
122 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
123 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
125 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE))
126 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * CFS_PAGE_SIZE"
128 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
129 # error "PTLRPC_MAX_BRW_SIZE too big"
131 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
132 # error "PTLRPC_MAX_BRW_PAGES too big"
134 #endif /* __KERNEL__ */
136 #define PTLRPC_NTHRS_INIT 2
141 * Constants determine how memory is used to buffer incoming service requests.
143 * ?_NBUFS # buffers to allocate when growing the pool
144 * ?_BUFSIZE # bytes in a single request buffer
145 * ?_MAXREQSIZE # maximum request service will receive
147 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
148 * of ?_NBUFS is added to the pool.
150 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
151 * considered full when less than ?_MAXREQSIZE is left in them.
156 * Constants determine how threads are created for ptlrpc service.
158 * ?_NTHRS_INIT # threads to create for each service partition on
159 * initializing. If it's non-affinity service and
160 * there is only one partition, it's the overall #
161 * threads for the service while initializing.
162 * ?_NTHRS_BASE # threads should be created at least for each
163 * ptlrpc partition to keep the service healthy.
164 * It's the low-water mark of threads upper-limit
165 * for each partition.
166 * ?_THR_FACTOR # threads can be added on threads upper-limit for
167 * each CPU core. This factor is only for reference,
168 * we might decrease value of factor if number of cores
169 * per CPT is above a limit.
170 * ?_NTHRS_MAX # overall threads can be created for a service,
171 * it's a soft limit because if service is running
172 * on machine with hundreds of cores and tens of
173 * CPU partitions, we need to guarantee each partition
174 * has ?_NTHRS_BASE threads, which means total threads
175 * will be ?_NTHRS_BASE * number_of_cpts which can
176 * exceed ?_NTHRS_MAX.
180 * #define MDS_NTHRS_INIT 2
181 * #define MDS_NTHRS_BASE 64
182 * #define MDS_NTHRS_FACTOR 8
183 * #define MDS_NTHRS_MAX 1024
186 * ---------------------------------------------------------------------
187 * Server(A) has 16 cores, user configured it to 4 partitions so each
188 * partition has 4 cores, then actual number of service threads on each
190 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
192 * Total number of threads for the service is:
193 * 96 * partitions(4) = 384
196 * ---------------------------------------------------------------------
197 * Server(B) has 32 cores, user configured it to 4 partitions so each
198 * partition has 8 cores, then actual number of service threads on each
200 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
202 * Total number of threads for the service is:
203 * 128 * partitions(4) = 512
206 * ---------------------------------------------------------------------
207 * Server(B) has 96 cores, user configured it to 8 partitions so each
208 * partition has 12 cores, then actual number of service threads on each
210 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
212 * Total number of threads for the service is:
213 * 160 * partitions(8) = 1280
215 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
216 * as upper limit of threads number for each partition:
217 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
220 * ---------------------------------------------------------------------
221 * Server(C) have a thousand of cores and user configured it to 32 partitions
222 * MDS_NTHRS_BASE(64) * 32 = 2048
224 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
225 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
226 * to keep service healthy, so total number of threads will just be 2048.
228 * NB: we don't suggest to choose server with that many cores because backend
229 * filesystem itself, buffer cache, or underlying network stack might
230 * have some SMP scalability issues at that large scale.
232 * If user already has a fat machine with hundreds or thousands of cores,
233 * there are two choices for configuration:
234 * a) create CPU table from subset of all CPUs and run Lustre on
236 * b) bind service threads on a few partitions, see modparameters of
237 * MDS and OSS for details
239 * NB: these calculations (and examples below) are simplified to help
240 * understanding, the real implementation is a little more complex,
241 * please see ptlrpc_server_nthreads_check() for details.
246 * LDLM threads constants:
248 * Given 8 as factor and 24 as base threads number
251 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
254 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
255 * threads for each partition and total threads number will be 112.
258 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
259 * threads for each partition to keep service healthy, so total threads
260 * number should be 24 * 8 = 192.
262 * So with these constants, threads number wil be at the similar level
263 * of old versions, unless target machine has over a hundred cores
265 #define LDLM_THR_FACTOR 8
266 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
267 #define LDLM_NTHRS_BASE 24
268 #define LDLM_NTHRS_MAX (cfs_num_online_cpus() == 1 ? 64 : 128)
270 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
271 #define LDLM_NBUFS (64 * cfs_num_online_cpus())
272 #define LDLM_BUFSIZE (8 * 1024)
273 #define LDLM_MAXREQSIZE (5 * 1024)
274 #define LDLM_MAXREPSIZE (1024)
277 * MDS threads constants:
279 * Please see examples in "Thread Constants", MDS threads number will be at
280 * the comparable level of old versions, unless the server has many cores.
282 #ifndef MDS_MAX_THREADS
283 #define MDS_MAX_THREADS 1024
284 #define MDS_MAX_OTHR_THREADS 256
286 #else /* MDS_MAX_THREADS */
287 #if MDS_MAX_THREADS < PTLRPC_NTHRS_INIT
288 #undef MDS_MAX_THREADS
289 #define MDS_MAX_THREADS PTLRPC_NTHRS_INIT
291 #define MDS_MAX_OTHR_THREADS max(PTLRPC_NTHRS_INIT, MDS_MAX_THREADS / 2)
294 /* default service */
295 #define MDS_THR_FACTOR 8
296 #define MDS_NTHRS_INIT PTLRPC_NTHRS_INIT
297 #define MDS_NTHRS_MAX MDS_MAX_THREADS
298 #define MDS_NTHRS_BASE min(64, MDS_NTHRS_MAX)
300 /* read-page service */
301 #define MDS_RDPG_THR_FACTOR 4
302 #define MDS_RDPG_NTHRS_INIT PTLRPC_NTHRS_INIT
303 #define MDS_RDPG_NTHRS_MAX MDS_MAX_OTHR_THREADS
304 #define MDS_RDPG_NTHRS_BASE min(48, MDS_RDPG_NTHRS_MAX)
306 /* these should be removed when we remove setattr service in the future */
307 #define MDS_SETA_THR_FACTOR 4
308 #define MDS_SETA_NTHRS_INIT PTLRPC_NTHRS_INIT
309 #define MDS_SETA_NTHRS_MAX MDS_MAX_OTHR_THREADS
310 #define MDS_SETA_NTHRS_BASE min(48, MDS_SETA_NTHRS_MAX)
312 /* non-affinity threads */
313 #define MDS_OTHR_NTHRS_INIT PTLRPC_NTHRS_INIT
314 #define MDS_OTHR_NTHRS_MAX MDS_MAX_OTHR_THREADS
316 #define MDS_NBUFS (64 * cfs_num_online_cpus())
318 * Assume file name length = FNAME_MAX = 256 (true for ext3).
319 * path name length = PATH_MAX = 4096
320 * LOV MD size max = EA_MAX = 48000 (2000 stripes)
321 * symlink: FNAME_MAX + PATH_MAX <- largest
322 * link: FNAME_MAX + PATH_MAX (mds_rec_link < mds_rec_create)
323 * rename: FNAME_MAX + FNAME_MAX
324 * open: FNAME_MAX + EA_MAX
326 * MDS_MAXREQSIZE ~= 4736 bytes =
327 * lustre_msg + ldlm_request + mdt_body + mds_rec_create + FNAME_MAX + PATH_MAX
328 * MDS_MAXREPSIZE ~= 8300 bytes = lustre_msg + llog_header
329 * or, for mds_close() and mds_reint_unlink() on a many-OST filesystem:
330 * = 9210 bytes = lustre_msg + mdt_body + 160 * (easize + cookiesize)
332 * Realistic size is about 512 bytes (20 character name + 128 char symlink),
333 * except in the open case where there are a large number of OSTs in a LOV.
335 #define MDS_MAXREPSIZE max(10 * 1024, 362 + LOV_MAX_STRIPE_COUNT * 56)
336 #define MDS_MAXREQSIZE MDS_MAXREPSIZE
338 /** MDS_BUFSIZE = max_reqsize + max sptlrpc payload size */
339 #define MDS_BUFSIZE (MDS_MAXREQSIZE + 1024)
341 /** FLD_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc */
342 #define FLD_MAXREQSIZE (160)
344 /** FLD_MAXREPSIZE == lustre_msg + ptlrpc_body */
345 #define FLD_MAXREPSIZE (152)
348 * SEQ_MAXREQSIZE == lustre_msg + __u32 padding + ptlrpc_body + opc + lu_range +
350 #define SEQ_MAXREQSIZE (160)
352 /** SEQ_MAXREPSIZE == lustre_msg + ptlrpc_body + lu_range */
353 #define SEQ_MAXREPSIZE (152)
355 /** MGS threads must be >= 3, see bug 22458 comment #28 */
356 #define MGS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
357 #define MGS_NTHRS_MAX 32
359 #define MGS_NBUFS (64 * cfs_num_online_cpus())
360 #define MGS_BUFSIZE (8 * 1024)
361 #define MGS_MAXREQSIZE (7 * 1024)
362 #define MGS_MAXREPSIZE (9 * 1024)
365 * OSS threads constants:
367 * Given 8 as factor and 64 as base threads number
370 * On 8-core server configured to 2 partitions, we will have
371 * 64 + 8 * 4 = 96 threads for each partition, 192 total threads.
374 * On 32-core machine configured to 4 partitions, we will have
375 * 64 + 8 * 8 = 112 threads for each partition, so total threads number
376 * will be 112 * 4 = 448.
379 * On 64-core machine configured to 4 partitions, we will have
380 * 64 + 16 * 8 = 192 threads for each partition, so total threads number
381 * will be 192 * 4 = 768 which is above limit OSS_NTHRS_MAX(512), so we
382 * cut off the value to OSS_NTHRS_MAX(512) / 4 which is 128 threads
383 * for each partition.
385 * So we can see that with these constants, threads number wil be at the
386 * similar level of old versions, unless the server has many cores.
388 /* depress threads factor for VM with small memory size */
389 #define OSS_THR_FACTOR min_t(int, 8, \
390 CFS_NUM_CACHEPAGES >> (28 - CFS_PAGE_SHIFT))
391 #define OSS_NTHRS_INIT (PTLRPC_NTHRS_INIT + 1)
392 #define OSS_NTHRS_BASE 64
393 #define OSS_NTHRS_MAX 512
395 /* threads for handling "create" request */
396 #define OSS_CR_THR_FACTOR 1
397 #define OSS_CR_NTHRS_INIT PTLRPC_NTHRS_INIT
398 #define OSS_CR_NTHRS_BASE 8
399 #define OSS_CR_NTHRS_MAX 64
403 * lustre_msg + obdo + obd_ioobj + DT_MAX_BRW_PAGES * niobuf_remote
405 * - single object with 16 pages is 512 bytes
406 * - OST_MAXREQSIZE must be at least 1 page of cookies plus some spillover
407 * - Must be a multiple of 1024
409 #define _OST_MAXREQSIZE_SUM (sizeof(struct lustre_msg) + sizeof(struct obdo) + \
410 sizeof(struct obd_ioobj) + DT_MAX_BRW_PAGES * \
411 sizeof(struct niobuf_remote))
412 #define OST_MAXREQSIZE (((_OST_MAXREQSIZE_SUM - 1) | (1024 - 1)) + 1)
414 #define OST_MAXREPSIZE (9 * 1024)
416 #define OST_NBUFS (64 * cfs_num_online_cpus())
417 #define OST_BUFSIZE (OST_MAXREQSIZE + 1024)
419 /* Macro to hide a typecast. */
420 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
423 * Structure to single define portal connection.
425 struct ptlrpc_connection {
426 /** linkage for connections hash table */
427 cfs_hlist_node_t c_hash;
428 /** Our own lnet nid for this connection */
430 /** Remote side nid for this connection */
431 lnet_process_id_t c_peer;
432 /** UUID of the other side */
433 struct obd_uuid c_remote_uuid;
434 /** reference counter for this connection */
435 cfs_atomic_t c_refcount;
438 /** Client definition for PortalRPC */
439 struct ptlrpc_client {
440 /** What lnet portal does this client send messages to by default */
441 __u32 cli_request_portal;
442 /** What portal do we expect replies on */
443 __u32 cli_reply_portal;
444 /** Name of the client */
448 /** state flags of requests */
449 /* XXX only ones left are those used by the bulk descs as well! */
450 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
451 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
453 #define REQ_MAX_ACK_LOCKS 8
455 union ptlrpc_async_args {
457 * Scratchpad for passing args to completion interpreter. Users
458 * cast to the struct of their choosing, and CLASSERT that this is
459 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
460 * a pointer to it here. The pointer_arg ensures this struct is at
461 * least big enough for that.
463 void *pointer_arg[11];
467 struct ptlrpc_request_set;
468 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
469 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
472 * Definition of request set structure.
473 * Request set is a list of requests (not necessary to the same target) that
474 * once populated with RPCs could be sent in parallel.
475 * There are two kinds of request sets. General purpose and with dedicated
476 * serving thread. Example of the latter is ptlrpcd set.
477 * For general purpose sets once request set started sending it is impossible
478 * to add new requests to such set.
479 * Provides a way to call "completion callbacks" when all requests in the set
482 struct ptlrpc_request_set {
483 cfs_atomic_t set_refcount;
484 /** number of in queue requests */
485 cfs_atomic_t set_new_count;
486 /** number of uncompleted requests */
487 cfs_atomic_t set_remaining;
488 /** wait queue to wait on for request events */
489 cfs_waitq_t set_waitq;
490 cfs_waitq_t *set_wakeup_ptr;
491 /** List of requests in the set */
492 cfs_list_t set_requests;
494 * List of completion callbacks to be called when the set is completed
495 * This is only used if \a set_interpret is NULL.
496 * Links struct ptlrpc_set_cbdata.
498 cfs_list_t set_cblist;
499 /** Completion callback, if only one. */
500 set_interpreter_func set_interpret;
501 /** opaq argument passed to completion \a set_interpret callback. */
504 * Lock for \a set_new_requests manipulations
505 * locked so that any old caller can communicate requests to
506 * the set holder who can then fold them into the lock-free set
508 spinlock_t set_new_req_lock;
509 /** List of new yet unsent requests. Only used with ptlrpcd now. */
510 cfs_list_t set_new_requests;
512 /** rq_status of requests that have been freed already */
514 /** Additional fields used by the flow control extension */
515 /** Maximum number of RPCs in flight */
516 int set_max_inflight;
517 /** Callback function used to generate RPCs */
518 set_producer_func set_producer;
519 /** opaq argument passed to the producer callback */
520 void *set_producer_arg;
524 * Description of a single ptrlrpc_set callback
526 struct ptlrpc_set_cbdata {
527 /** List linkage item */
529 /** Pointer to interpreting function */
530 set_interpreter_func psc_interpret;
531 /** Opaq argument to pass to the callback */
535 struct ptlrpc_bulk_desc;
536 struct ptlrpc_service_part;
537 struct ptlrpc_service;
540 * ptlrpc callback & work item stuff
542 struct ptlrpc_cb_id {
543 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
544 void *cbid_arg; /* additional arg */
547 /** Maximum number of locks to fit into reply state */
548 #define RS_MAX_LOCKS 8
552 * Structure to define reply state on the server
553 * Reply state holds various reply message information. Also for "difficult"
554 * replies (rep-ack case) we store the state after sending reply and wait
555 * for the client to acknowledge the reception. In these cases locks could be
556 * added to the state for replay/failover consistency guarantees.
558 struct ptlrpc_reply_state {
559 /** Callback description */
560 struct ptlrpc_cb_id rs_cb_id;
561 /** Linkage for list of all reply states in a system */
563 /** Linkage for list of all reply states on same export */
564 cfs_list_t rs_exp_list;
565 /** Linkage for list of all reply states for same obd */
566 cfs_list_t rs_obd_list;
568 cfs_list_t rs_debug_list;
570 /** A spinlock to protect the reply state flags */
572 /** Reply state flags */
573 unsigned long rs_difficult:1; /* ACK/commit stuff */
574 unsigned long rs_no_ack:1; /* no ACK, even for
575 difficult requests */
576 unsigned long rs_scheduled:1; /* being handled? */
577 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
578 unsigned long rs_handled:1; /* been handled yet? */
579 unsigned long rs_on_net:1; /* reply_out_callback pending? */
580 unsigned long rs_prealloc:1; /* rs from prealloc list */
581 unsigned long rs_committed:1;/* the transaction was committed
582 and the rs was dispatched
583 by ptlrpc_commit_replies */
584 /** Size of the state */
588 /** Transaction number */
592 struct obd_export *rs_export;
593 struct ptlrpc_service_part *rs_svcpt;
594 /** Lnet metadata handle for the reply */
595 lnet_handle_md_t rs_md_h;
596 cfs_atomic_t rs_refcount;
598 /** Context for the sevice thread */
599 struct ptlrpc_svc_ctx *rs_svc_ctx;
600 /** Reply buffer (actually sent to the client), encoded if needed */
601 struct lustre_msg *rs_repbuf; /* wrapper */
602 /** Size of the reply buffer */
603 int rs_repbuf_len; /* wrapper buf length */
604 /** Size of the reply message */
605 int rs_repdata_len; /* wrapper msg length */
607 * Actual reply message. Its content is encrupted (if needed) to
608 * produce reply buffer for actual sending. In simple case
609 * of no network encryption we jus set \a rs_repbuf to \a rs_msg
611 struct lustre_msg *rs_msg; /* reply message */
613 /** Number of locks awaiting client ACK */
615 /** Handles of locks awaiting client reply ACK */
616 struct lustre_handle rs_locks[RS_MAX_LOCKS];
617 /** Lock modes of locks in \a rs_locks */
618 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
621 struct ptlrpc_thread;
625 RQ_PHASE_NEW = 0xebc0de00,
626 RQ_PHASE_RPC = 0xebc0de01,
627 RQ_PHASE_BULK = 0xebc0de02,
628 RQ_PHASE_INTERPRET = 0xebc0de03,
629 RQ_PHASE_COMPLETE = 0xebc0de04,
630 RQ_PHASE_UNREGISTERING = 0xebc0de05,
631 RQ_PHASE_UNDEFINED = 0xebc0de06
634 /** Type of request interpreter call-back */
635 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
636 struct ptlrpc_request *req,
640 * Definition of request pool structure.
641 * The pool is used to store empty preallocated requests for the case
642 * when we would actually need to send something without performing
643 * any allocations (to avoid e.g. OOM).
645 struct ptlrpc_request_pool {
646 /** Locks the list */
648 /** list of ptlrpc_request structs */
649 cfs_list_t prp_req_list;
650 /** Maximum message size that would fit into a rquest from this pool */
652 /** Function to allocate more requests for this pool */
653 void (*prp_populate)(struct ptlrpc_request_pool *, int);
662 * \defgroup nrs Network Request Scheduler
665 struct ptlrpc_nrs_policy;
666 struct ptlrpc_nrs_resource;
667 struct ptlrpc_nrs_request;
670 * NRS control operations.
672 * These are common for all policies.
674 enum ptlrpc_nrs_ctl {
676 * Activate the policy.
678 PTLRPC_NRS_CTL_START,
680 * Reserved for multiple primary policies, which may be a possibility
685 * Recycle resources for inactive policies.
687 PTLRPC_NRS_CTL_SHRINK,
689 * Not a valid opcode.
691 PTLRPC_NRS_CTL_INVALID,
693 * Policies can start using opcodes from this value and onwards for
694 * their own purposes; the assigned value itself is arbitrary.
696 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
700 * NRS policy operations.
702 * These determine the behaviour of a policy, and are called in response to
705 struct ptlrpc_nrs_pol_ops {
707 * Called during policy registration; this operation is optional.
709 * \param[in] policy The policy being initialized
711 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
713 * Called during policy unregistration; this operation is optional.
715 * \param[in] policy The policy being unregistered/finalized
717 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
719 * Called when activating a policy via lprocfs; policies allocate and
720 * initialize their resources here; this operation is optional.
722 * \param[in] policy The policy being started
724 * \see nrs_policy_start_locked()
726 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
728 * Called when deactivating a policy via lprocfs; policies deallocate
729 * their resources here; this operation is optional
731 * \param[in] policy The policy being stopped
733 * \see nrs_policy_stop_final()
735 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
737 * Used for policy-specific operations; i.e. not generic ones like
738 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
739 * to an ioctl; this operation is optional.
741 * \param[in] policy The policy carrying out operation \a opc
742 * \param[in] opc The command operation being carried out
743 * \param[in,out] arg An generic buffer for communication between the
744 * user and the control operation
749 * \see ptlrpc_nrs_policy_control()
751 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
752 enum ptlrpc_nrs_ctl opc, void *arg);
755 * Called when obtaining references to the resources of the resource
756 * hierarchy for a request that has arrived for handling at the PTLRPC
757 * service. Policies should return -ve for requests they do not wish
758 * to handle. This operation is mandatory.
760 * \param[in] policy The policy we're getting resources for.
761 * \param[in] nrq The request we are getting resources for.
762 * \param[in] parent The parent resource of the resource being
763 * requested; set to NULL if none.
764 * \param[out] resp The resource is to be returned here; the
765 * fallback policy in an NRS head should
766 * \e always return a non-NULL pointer value.
767 * \param[in] moving_req When set, signifies that this is an attempt
768 * to obtain resources for a request being moved
769 * to the high-priority NRS head by
770 * ldlm_lock_reorder_req().
771 * This implies two things:
772 * 1. We are under obd_export::exp_rpc_lock and
773 * so should not sleep.
774 * 2. We should not perform non-idempotent or can
775 * skip performing idempotent operations that
776 * were carried out when resources were first
777 * taken for the request when it was initialized
778 * in ptlrpc_nrs_req_initialize().
780 * \retval 0, +ve The level of the returned resource in the resource
781 * hierarchy; currently only 0 (for a non-leaf resource)
782 * and 1 (for a leaf resource) are supported by the
786 * \see ptlrpc_nrs_req_initialize()
787 * \see ptlrpc_nrs_hpreq_add_nolock()
788 * \see ptlrpc_nrs_req_hp_move()
790 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
791 struct ptlrpc_nrs_request *nrq,
792 struct ptlrpc_nrs_resource *parent,
793 struct ptlrpc_nrs_resource **resp,
796 * Called when releasing references taken for resources in the resource
797 * hierarchy for the request; this operation is optional.
799 * \param[in] policy The policy the resource belongs to
800 * \param[in] res The resource to be freed
802 * \see ptlrpc_nrs_req_finalize()
803 * \see ptlrpc_nrs_hpreq_add_nolock()
804 * \see ptlrpc_nrs_req_hp_move()
806 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
807 struct ptlrpc_nrs_resource *res);
810 * Obtain a request for handling from the policy via polling; this
811 * operation is mandatory.
813 * \param[in] policy The policy to poll
815 * \retval NULL No erquest available for handling
816 * \retval valid-pointer The request polled for handling
818 * \see ptlrpc_nrs_req_poll_nolock()
820 struct ptlrpc_nrs_request *
821 (*op_req_poll) (struct ptlrpc_nrs_policy *policy);
823 * Called when attempting to add a request to a policy for later
824 * handling; this operation is mandatory.
826 * \param[in] policy The policy on which to enqueue \a nrq
827 * \param[in] nrq The request to enqueue
832 * \see ptlrpc_nrs_req_add_nolock()
834 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
835 struct ptlrpc_nrs_request *nrq);
837 * Removes a request from the policy's set of pending requests. Normally
838 * called after a request has been polled successfully from the policy
839 * for handling; this operation is mandatory.
841 * \param[in] policy The policy the request \a nrq belongs to
842 * \param[in] nrq The request to dequeue
844 * \see ptlrpc_nrs_req_del_nolock()
846 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
847 struct ptlrpc_nrs_request *nrq);
849 * Called before carrying out the request; should not block. Could be
850 * used for job/resource control; this operation is optional.
852 * \param[in] policy The policy which is starting to handle request
854 * \param[in] nrq The request
856 * \pre spin_is_locked(&svcpt->scp_req_lock)
858 * \see ptlrpc_nrs_req_start_nolock()
860 void (*op_req_start) (struct ptlrpc_nrs_policy *policy,
861 struct ptlrpc_nrs_request *nrq);
863 * Called after the request being carried out. Could be used for
864 * job/resource control; this operation is optional.
866 * \param[in] policy The policy which is stopping to handle request
868 * \param[in] nrq The request
870 * \pre spin_is_locked(&svcpt->scp_req_lock)
872 * \see ptlrpc_nrs_req_stop_nolock()
874 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
875 struct ptlrpc_nrs_request *nrq);
877 * Registers the policy's lprocfs interface with a PTLRPC service.
879 * \param[in] svc The service
884 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
886 * Unegisters the policy's lprocfs interface with a PTLRPC service.
888 * \param[in] svc The service
890 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
896 enum nrs_policy_flags {
898 * Fallback policy, use this flag only on a single supported policy per
899 * service. Do not use this flag for policies registering using
900 * ptlrpc_nrs_policy_register() (i.e. ones that are not in
901 * \e nrs_pols_builtin).
903 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
905 * Start policy immediately after registering.
907 PTLRPC_NRS_FL_REG_START = (1 << 1),
909 * This is a polciy registering externally with NRS core, via
910 * ptlrpc_nrs_policy_register(), (i.e. one that is not in
911 * \e nrs_pols_builtin. Used to avoid ptlrpc_nrs_policy_register()
912 * racing with a policy start operation issued by the user via lprocfs.
914 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
920 * Denotes whether an NRS instance is for handling normal or high-priority
921 * RPCs, or whether an operation pertains to one or both of the NRS instances
924 enum ptlrpc_nrs_queue_type {
925 PTLRPC_NRS_QUEUE_REG,
927 PTLRPC_NRS_QUEUE_BOTH,
933 * A PTLRPC service has at least one NRS head instance for handling normal
934 * priority RPCs, and may optionally have a second NRS head instance for
935 * handling high-priority RPCs. Each NRS head maintains a list of available
936 * policies, of which one and only one policy is acting as the fallback policy,
937 * and optionally a different policy may be acting as the primary policy. For
938 * all RPCs handled by this NRS head instance, NRS core will first attempt to
939 * enqueue the RPC using the primary policy (if any). The fallback policy is
940 * used in the following cases:
941 * - when there was no primary policy in the
942 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
944 * - when the primary policy that was at the
945 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
946 * RPC was initialized, denoted it did not wish, or for some other reason was
947 * not able to handle the request, by returning a non-valid NRS resource
949 * - when the primary policy that was at the
950 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
951 * RPC was initialized, fails later during the request enqueueing stage.
953 * \see nrs_resource_get_safe()
954 * \see nrs_request_enqueue()
958 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
960 * Linkage into nrs_core_heads_list
962 cfs_list_t nrs_heads;
964 * List of registered policies
966 cfs_list_t nrs_policy_list;
968 * List of policies with queued requests. Policies that have any
969 * outstanding requests are queued here, and this list is queried
970 * in a round-robin manner from NRS core when obtaining a request
971 * for handling. This ensures that requests from policies that at some
972 * point transition away from the
973 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
975 cfs_list_t nrs_policy_queued;
977 * Service partition for this NRS head
979 struct ptlrpc_service_part *nrs_svcpt;
981 * Primary policy, which is the preferred policy for handling RPCs
983 struct ptlrpc_nrs_policy *nrs_policy_primary;
985 * Fallback policy, which is the backup policy for handling RPCs
987 struct ptlrpc_nrs_policy *nrs_policy_fallback;
989 * This NRS head handles either HP or regular requests
991 enum ptlrpc_nrs_queue_type nrs_queue_type;
993 * # queued requests from all policies in this NRS head
995 unsigned long nrs_req_queued;
997 * # scheduled requests from all policies in this NRS head
999 unsigned long nrs_req_started;
1001 * # policies on this NRS
1002 * TODO: Can we avoid having this?
1004 unsigned nrs_num_pols;
1006 * This NRS head is in progress of starting a policy
1008 unsigned nrs_policy_starting:1;
1010 * In progress of shutting down the whole NRS head; used during
1013 unsigned nrs_stopping:1;
1016 #define NRS_POL_NAME_MAX 16
1019 * NRS policy registering descriptor
1021 * Is used to hold a description of a policy that can be passed to NRS core in
1022 * order to register the policy with NRS heads in different PTLRPC services.
1024 struct ptlrpc_nrs_pol_desc {
1026 * Human-readable policy name
1028 char pd_name[NRS_POL_NAME_MAX];
1030 * NRS operations for this policy
1032 struct ptlrpc_nrs_pol_ops *pd_ops;
1034 * Service Compatibility function; this determines whether a policy is
1035 * adequate for handling RPCs of a particular PTLRPC service.
1037 * XXX:This should give the same result during policy
1038 * registration and unregistration, and for all partitions of a
1039 * service; so the result should not depend on temporal service
1040 * or other properties, that may influence the result.
1042 bool (*pd_compat) (struct ptlrpc_service *svc,
1043 const struct ptlrpc_nrs_pol_desc *desc);
1045 * Optionally set for policies that support a single ptlrpc service,
1046 * i.e. ones that have \a pd_compat set to nrs_policy_compat_one()
1048 char *pd_compat_svc_name;
1050 * Bitmask of nrs_policy_flags
1054 * Link into nrs_core::nrs_policies
1062 * Policies transition from one state to the other during their lifetime
1064 enum ptlrpc_nrs_pol_state {
1066 * Not a valid policy state.
1068 NRS_POL_STATE_INVALID,
1070 * For now, this state is used exclusively for policies that register
1071 * externally to NRS core, i.e. ones that do so via
1072 * ptlrpc_nrs_policy_register() and are not part of nrs_pols_builtin;
1073 * it is used to prevent a race condition between the policy registering
1074 * with more than one service partition while service is operational,
1075 * and the user starting the policy via lprocfs.
1077 * \see nrs_pol_make_avail()
1079 NRS_POL_STATE_UNAVAIL,
1081 * Policies are at this state either at the start of their life, or
1082 * transition here when the user selects a different policy to act
1083 * as the primary one.
1085 NRS_POL_STATE_STOPPED,
1087 * Policy is progress of stopping
1089 NRS_POL_STATE_STOPPING,
1091 * Policy is in progress of starting
1093 NRS_POL_STATE_STARTING,
1095 * A policy is in this state in two cases:
1096 * - it is the fallback policy, which is always in this state.
1097 * - it has been activated by the user; i.e. it is the primary policy,
1099 NRS_POL_STATE_STARTED,
1103 * NRS policy information
1105 * Used for obtaining information for the status of a policy via lprocfs
1107 struct ptlrpc_nrs_pol_info {
1111 char pi_name[NRS_POL_NAME_MAX];
1113 * Current policy state
1115 enum ptlrpc_nrs_pol_state pi_state;
1117 * # RPCs enqueued for later dispatching by the policy
1121 * # RPCs started for dispatch by the policy
1123 long pi_req_started;
1125 * Is this a fallback policy?
1127 unsigned pi_fallback:1;
1133 * There is one instance of this for each policy in each NRS head of each
1134 * PTLRPC service partition.
1136 struct ptlrpc_nrs_policy {
1138 * Linkage into the NRS head's list of policies,
1139 * ptlrpc_nrs:nrs_policy_list
1141 cfs_list_t pol_list;
1143 * Linkage into the NRS head's list of policies with enqueued
1144 * requests ptlrpc_nrs:nrs_policy_queued
1146 cfs_list_t pol_list_queued;
1148 * Current state of this policy
1150 enum ptlrpc_nrs_pol_state pol_state;
1152 * Bitmask of nrs_policy_flags
1156 * # RPCs enqueued for later dispatching by the policy
1158 long pol_req_queued;
1160 * # RPCs started for dispatch by the policy
1162 long pol_req_started;
1164 * Usage Reference count taken on the policy instance
1168 * The NRS head this policy has been created at
1170 struct ptlrpc_nrs *pol_nrs;
1172 * NRS operations for this policy; points to ptlrpc_nrs_pol_desc::pd_ops
1174 struct ptlrpc_nrs_pol_ops *pol_ops;
1176 * Private policy data; varies by policy type
1180 * Human-readable policy name; point to ptlrpc_nrs_pol_desc::pd_name
1188 * Resources are embedded into two types of NRS entities:
1189 * - Inside NRS policies, in the policy's private data in
1190 * ptlrpc_nrs_policy::pol_private
1191 * - In objects that act as prime-level scheduling entities in different NRS
1192 * policies; e.g. on a policy that performs round robin or similar order
1193 * scheduling across client NIDs, there would be one NRS resource per unique
1194 * client NID. On a policy which performs round robin scheduling across
1195 * backend filesystem objects, there would be one resource associated with
1196 * each of the backend filesystem objects partaking in the scheduling
1197 * performed by the policy.
1199 * NRS resources share a parent-child relationship, in which resources embedded
1200 * in policy instances are the parent entities, with all scheduling entities
1201 * a policy schedules across being the children, thus forming a simple resource
1202 * hierarchy. This hierarchy may be extended with one or more levels in the
1203 * future if the ability to have more than one primary policy is added.
1205 * Upon request initialization, references to the then active NRS policies are
1206 * taken and used to later handle the dispatching of the request with one of
1209 * \see nrs_resource_get_safe()
1210 * \see ptlrpc_nrs_req_add()
1212 struct ptlrpc_nrs_resource {
1214 * This NRS resource's parent; is NULL for resources embedded in NRS
1215 * policy instances; i.e. those are top-level ones.
1217 struct ptlrpc_nrs_resource *res_parent;
1219 * The policy associated with this resource.
1221 struct ptlrpc_nrs_policy *res_policy;
1234 * This policy is a logical wrapper around previous, non-NRS functionality.
1235 * It dispatches RPCs in the same order as they arrive from the network. This
1236 * policy is currently used as the fallback policy, and the only enabled policy
1237 * on all NRS heads of all PTLRPC service partitions.
1242 * Private data structure for the FIFO policy
1244 struct nrs_fifo_head {
1246 * Resource object for policy instance.
1248 struct ptlrpc_nrs_resource fh_res;
1250 * List of queued requests.
1254 * For debugging purposes.
1259 struct nrs_fifo_req {
1260 /** request header, must be the first member of structure */
1270 * Instances of this object exist embedded within ptlrpc_request; the main
1271 * purpose of this object is to hold references to the request's resources
1272 * for the lifetime of the request, and to hold properties that policies use
1273 * use for determining the request's scheduling priority.
1275 struct ptlrpc_nrs_request {
1277 * The request's resource hierarchy.
1279 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1281 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1282 * policy that was used to enqueue the request.
1284 * \see nrs_request_enqueue()
1286 unsigned nr_res_idx;
1287 unsigned nr_initialized:1;
1288 unsigned nr_enqueued:1;
1289 unsigned nr_dequeued:1;
1290 unsigned nr_started:1;
1291 unsigned nr_finalized:1;
1292 cfs_binheap_node_t nr_node;
1295 * Policy-specific fields, used for determining a request's scheduling
1296 * priority, and other supporting functionality.
1300 * Fields for the FIFO policy
1302 struct nrs_fifo_req fifo;
1305 * Externally-registering policies may want to use this to allocate
1306 * their own request properties.
1314 * Basic request prioritization operations structure.
1315 * The whole idea is centered around locks and RPCs that might affect locks.
1316 * When a lock is contended we try to give priority to RPCs that might lead
1317 * to fastest release of that lock.
1318 * Currently only implemented for OSTs only in a way that makes all
1319 * IO and truncate RPCs that are coming from a locked region where a lock is
1320 * contended a priority over other requests.
1322 struct ptlrpc_hpreq_ops {
1324 * Check if the lock handle of the given lock is the same as
1325 * taken from the request.
1327 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1329 * Check if the request is a high priority one.
1331 int (*hpreq_check)(struct ptlrpc_request *);
1333 * Called after the request has been handled.
1335 void (*hpreq_fini)(struct ptlrpc_request *);
1339 * Represents remote procedure call.
1341 * This is a staple structure used by everybody wanting to send a request
1344 struct ptlrpc_request {
1345 /* Request type: one of PTL_RPC_MSG_* */
1347 /** Result of request processing */
1350 * Linkage item through which this request is included into
1351 * sending/delayed lists on client and into rqbd list on server
1355 * Server side list of incoming unserved requests sorted by arrival
1356 * time. Traversed from time to time to notice about to expire
1357 * requests and sent back "early replies" to clients to let them
1358 * know server is alive and well, just very busy to service their
1361 cfs_list_t rq_timed_list;
1362 /** server-side history, used for debuging purposes. */
1363 cfs_list_t rq_history_list;
1364 /** server-side per-export list */
1365 cfs_list_t rq_exp_list;
1366 /** server-side hp handlers */
1367 struct ptlrpc_hpreq_ops *rq_ops;
1369 /** initial thread servicing this request */
1370 struct ptlrpc_thread *rq_svc_thread;
1372 /** history sequence # */
1373 __u64 rq_history_seq;
1377 /** stub for NRS request */
1378 struct ptlrpc_nrs_request rq_nrq;
1380 /** the index of service's srv_at_array into which request is linked */
1382 /** Lock to protect request flags and some other important bits, like
1386 /** client-side flags are serialized by rq_lock */
1387 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1388 rq_timedout:1, rq_resend:1, rq_restart:1,
1390 * when ->rq_replay is set, request is kept by the client even
1391 * after server commits corresponding transaction. This is
1392 * used for operations that require sequence of multiple
1393 * requests to be replayed. The only example currently is file
1394 * open/close. When last request in such a sequence is
1395 * committed, ->rq_replay is cleared on all requests in the
1399 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1400 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1401 rq_early:1, rq_must_unlink:1,
1402 rq_memalloc:1, /* req originated from "kswapd" */
1403 /* server-side flags */
1404 rq_packed_final:1, /* packed final reply */
1405 rq_hp:1, /* high priority RPC */
1406 rq_at_linked:1, /* link into service's srv_at_array */
1407 rq_reply_truncate:1,
1409 /* whether the "rq_set" is a valid one */
1411 rq_generation_set:1,
1412 /* do not resend request on -EINPROGRESS */
1413 rq_no_retry_einprogress:1,
1414 /* allow the req to be sent if the import is in recovery
1418 unsigned int rq_nr_resend;
1420 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1421 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1422 cfs_atomic_t rq_refcount;/* client-side refcount for SENT race,
1423 server-side refcounf for multiple replies */
1425 /** Portal to which this request would be sent */
1426 short rq_request_portal; /* XXX FIXME bug 249 */
1427 /** Portal where to wait for reply and where reply would be sent */
1428 short rq_reply_portal; /* XXX FIXME bug 249 */
1432 * !rq_truncate : # reply bytes actually received,
1433 * rq_truncate : required repbuf_len for resend
1435 int rq_nob_received;
1436 /** Request length */
1440 /** Request message - what client sent */
1441 struct lustre_msg *rq_reqmsg;
1442 /** Reply message - server response */
1443 struct lustre_msg *rq_repmsg;
1444 /** Transaction number */
1449 * List item to for replay list. Not yet commited requests get linked
1451 * Also see \a rq_replay comment above.
1453 cfs_list_t rq_replay_list;
1456 * security and encryption data
1458 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1459 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1460 cfs_list_t rq_ctx_chain; /**< link to waited ctx */
1462 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1463 enum lustre_sec_part rq_sp_from;
1465 /* client/server security flags */
1467 rq_ctx_init:1, /* context initiation */
1468 rq_ctx_fini:1, /* context destroy */
1469 rq_bulk_read:1, /* request bulk read */
1470 rq_bulk_write:1, /* request bulk write */
1471 /* server authentication flags */
1472 rq_auth_gss:1, /* authenticated by gss */
1473 rq_auth_remote:1, /* authed as remote user */
1474 rq_auth_usr_root:1, /* authed as root */
1475 rq_auth_usr_mdt:1, /* authed as mdt */
1476 rq_auth_usr_ost:1, /* authed as ost */
1477 /* security tfm flags */
1480 /* doesn't expect reply FIXME */
1482 rq_pill_init:1; /* pill initialized */
1484 uid_t rq_auth_uid; /* authed uid */
1485 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1487 /* (server side), pointed directly into req buffer */
1488 struct ptlrpc_user_desc *rq_user_desc;
1490 /* various buffer pointers */
1491 struct lustre_msg *rq_reqbuf; /* req wrapper */
1492 char *rq_repbuf; /* rep buffer */
1493 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1494 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1495 int rq_reqbuf_len; /* req wrapper buf len */
1496 int rq_reqdata_len; /* req wrapper msg len */
1497 int rq_repbuf_len; /* rep buffer len */
1498 int rq_repdata_len; /* rep wrapper msg len */
1499 int rq_clrbuf_len; /* only in priv mode */
1500 int rq_clrdata_len; /* only in priv mode */
1502 /** early replies go to offset 0, regular replies go after that */
1503 unsigned int rq_reply_off;
1507 /** Fields that help to see if request and reply were swabbed or not */
1508 __u32 rq_req_swab_mask;
1509 __u32 rq_rep_swab_mask;
1511 /** What was import generation when this request was sent */
1512 int rq_import_generation;
1513 enum lustre_imp_state rq_send_state;
1515 /** how many early replies (for stats) */
1518 /** client+server request */
1519 lnet_handle_md_t rq_req_md_h;
1520 struct ptlrpc_cb_id rq_req_cbid;
1521 /** optional time limit for send attempts */
1522 cfs_duration_t rq_delay_limit;
1523 /** time request was first queued */
1524 cfs_time_t rq_queued_time;
1526 /* server-side... */
1527 /** request arrival time */
1528 struct timeval rq_arrival_time;
1529 /** separated reply state */
1530 struct ptlrpc_reply_state *rq_reply_state;
1531 /** incoming request buffer */
1532 struct ptlrpc_request_buffer_desc *rq_rqbd;
1534 /** client-only incoming reply */
1535 lnet_handle_md_t rq_reply_md_h;
1536 cfs_waitq_t rq_reply_waitq;
1537 struct ptlrpc_cb_id rq_reply_cbid;
1541 /** Peer description (the other side) */
1542 lnet_process_id_t rq_peer;
1543 /** Server-side, export on which request was received */
1544 struct obd_export *rq_export;
1545 /** Client side, import where request is being sent */
1546 struct obd_import *rq_import;
1548 /** Replay callback, called after request is replayed at recovery */
1549 void (*rq_replay_cb)(struct ptlrpc_request *);
1551 * Commit callback, called when request is committed and about to be
1554 void (*rq_commit_cb)(struct ptlrpc_request *);
1555 /** Opaq data for replay and commit callbacks. */
1558 /** For bulk requests on client only: bulk descriptor */
1559 struct ptlrpc_bulk_desc *rq_bulk;
1561 /** client outgoing req */
1563 * when request/reply sent (secs), or time when request should be sent
1566 /** time for request really sent out */
1567 time_t rq_real_sent;
1569 /** when request must finish. volatile
1570 * so that servers' early reply updates to the deadline aren't
1571 * kept in per-cpu cache */
1572 volatile time_t rq_deadline;
1573 /** when req reply unlink must finish. */
1574 time_t rq_reply_deadline;
1575 /** when req bulk unlink must finish. */
1576 time_t rq_bulk_deadline;
1578 * service time estimate (secs)
1579 * If the requestsis not served by this time, it is marked as timed out.
1583 /** Multi-rpc bits */
1584 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1585 cfs_waitq_t rq_set_waitq;
1586 /** Link item for request set lists */
1587 cfs_list_t rq_set_chain;
1588 /** Link back to the request set */
1589 struct ptlrpc_request_set *rq_set;
1590 /** Async completion handler, called when reply is received */
1591 ptlrpc_interpterer_t rq_interpret_reply;
1592 /** Async completion context */
1593 union ptlrpc_async_args rq_async_args;
1595 /** Pool if request is from preallocated list */
1596 struct ptlrpc_request_pool *rq_pool;
1598 struct lu_context rq_session;
1599 struct lu_context rq_recov_session;
1601 /** request format description */
1602 struct req_capsule rq_pill;
1606 * Call completion handler for rpc if any, return it's status or original
1607 * rc if there was no handler defined for this request.
1609 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1610 struct ptlrpc_request *req, int rc)
1612 if (req->rq_interpret_reply != NULL) {
1613 req->rq_status = req->rq_interpret_reply(env, req,
1614 &req->rq_async_args,
1616 return req->rq_status;
1624 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_desc *desc);
1625 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_desc *desc);
1626 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1627 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1628 struct ptlrpc_nrs_pol_info *info);
1631 * Can the request be moved from the regular NRS head to the high-priority NRS
1632 * head (of the same PTLRPC service partition), if any?
1634 * For a reliable result, this should be checked under svcpt->scp_req lock.
1637 ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1639 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1642 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1643 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1644 * to make sure it has not been scheduled yet (analogous to previous
1645 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1647 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1652 * Returns 1 if request buffer at offset \a index was already swabbed
1654 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1656 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1657 return req->rq_req_swab_mask & (1 << index);
1661 * Returns 1 if request reply buffer at offset \a index was already swabbed
1663 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1665 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1666 return req->rq_rep_swab_mask & (1 << index);
1670 * Returns 1 if request needs to be swabbed into local cpu byteorder
1672 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1674 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1678 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1680 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1682 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1686 * Mark request buffer at offset \a index that it was already swabbed
1688 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1690 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1691 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1692 req->rq_req_swab_mask |= 1 << index;
1696 * Mark request reply buffer at offset \a index that it was already swabbed
1698 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1700 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1701 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1702 req->rq_rep_swab_mask |= 1 << index;
1706 * Convert numerical request phase value \a phase into text string description
1708 static inline const char *
1709 ptlrpc_phase2str(enum rq_phase phase)
1718 case RQ_PHASE_INTERPRET:
1720 case RQ_PHASE_COMPLETE:
1722 case RQ_PHASE_UNREGISTERING:
1723 return "Unregistering";
1730 * Convert numerical request phase of the request \a req into text stringi
1733 static inline const char *
1734 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1736 return ptlrpc_phase2str(req->rq_phase);
1740 * Debugging functions and helpers to print request structure into debug log
1743 /* Spare the preprocessor, spoil the bugs. */
1744 #define FLAG(field, str) (field ? str : "")
1746 /** Convert bit flags into a string */
1747 #define DEBUG_REQ_FLAGS(req) \
1748 ptlrpc_rqphase2str(req), \
1749 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1750 FLAG(req->rq_err, "E"), \
1751 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1752 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1753 FLAG(req->rq_no_resend, "N"), \
1754 FLAG(req->rq_waiting, "W"), \
1755 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1756 FLAG(req->rq_committed, "M")
1758 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1760 void _debug_req(struct ptlrpc_request *req,
1761 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1762 __attribute__ ((format (printf, 3, 4)));
1765 * Helper that decides if we need to print request accordig to current debug
1768 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1770 CFS_CHECK_STACK(msgdata, mask, cdls); \
1772 if (((mask) & D_CANTMASK) != 0 || \
1773 ((libcfs_debug & (mask)) != 0 && \
1774 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1775 _debug_req((req), msgdata, fmt, ##a); \
1779 * This is the debug print function you need to use to print request sturucture
1780 * content into lustre debug log.
1781 * for most callers (level is a constant) this is resolved at compile time */
1782 #define DEBUG_REQ(level, req, fmt, args...) \
1784 if ((level) & (D_ERROR | D_WARNING)) { \
1785 static cfs_debug_limit_state_t cdls; \
1786 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1787 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1789 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1790 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1796 * Structure that defines a single page of a bulk transfer
1798 struct ptlrpc_bulk_page {
1799 /** Linkage to list of pages in a bulk */
1802 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1805 /** offset within a page */
1807 /** The page itself */
1808 struct page *bp_page;
1811 #define BULK_GET_SOURCE 0
1812 #define BULK_PUT_SINK 1
1813 #define BULK_GET_SINK 2
1814 #define BULK_PUT_SOURCE 3
1817 * Definition of bulk descriptor.
1818 * Bulks are special "Two phase" RPCs where initial request message
1819 * is sent first and it is followed bt a transfer (o receiving) of a large
1820 * amount of data to be settled into pages referenced from the bulk descriptors.
1821 * Bulks transfers (the actual data following the small requests) are done
1822 * on separate LNet portals.
1823 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1824 * Another user is readpage for MDT.
1826 struct ptlrpc_bulk_desc {
1827 /** completed with failure */
1828 unsigned long bd_failure:1;
1829 /** {put,get}{source,sink} */
1830 unsigned long bd_type:2;
1832 unsigned long bd_registered:1;
1833 /** For serialization with callback */
1835 /** Import generation when request for this bulk was sent */
1836 int bd_import_generation;
1837 /** LNet portal for this bulk */
1839 /** Server side - export this bulk created for */
1840 struct obd_export *bd_export;
1841 /** Client side - import this bulk was sent on */
1842 struct obd_import *bd_import;
1843 /** Back pointer to the request */
1844 struct ptlrpc_request *bd_req;
1845 cfs_waitq_t bd_waitq; /* server side only WQ */
1846 int bd_iov_count; /* # entries in bd_iov */
1847 int bd_max_iov; /* allocated size of bd_iov */
1848 int bd_nob; /* # bytes covered */
1849 int bd_nob_transferred; /* # bytes GOT/PUT */
1853 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1854 lnet_nid_t bd_sender; /* stash event::sender */
1855 int bd_md_count; /* # valid entries in bd_mds */
1856 int bd_md_max_brw; /* max entries in bd_mds */
1857 /** array of associated MDs */
1858 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1860 #if defined(__KERNEL__)
1862 * encrypt iov, size is either 0 or bd_iov_count.
1864 lnet_kiov_t *bd_enc_iov;
1866 lnet_kiov_t bd_iov[0];
1868 lnet_md_iovec_t bd_iov[0];
1873 SVC_STOPPED = 1 << 0,
1874 SVC_STOPPING = 1 << 1,
1875 SVC_STARTING = 1 << 2,
1876 SVC_RUNNING = 1 << 3,
1878 SVC_SIGNAL = 1 << 5,
1881 #define PTLRPC_THR_NAME_LEN 32
1883 * Definition of server service thread structure
1885 struct ptlrpc_thread {
1887 * List of active threads in svc->srv_threads
1891 * thread-private data (preallocated memory)
1896 * service thread index, from ptlrpc_start_threads
1900 * service thread pid
1904 * put watchdog in the structure per thread b=14840
1906 struct lc_watchdog *t_watchdog;
1908 * the svc this thread belonged to b=18582
1910 struct ptlrpc_service_part *t_svcpt;
1911 cfs_waitq_t t_ctl_waitq;
1912 struct lu_env *t_env;
1913 char t_name[PTLRPC_THR_NAME_LEN];
1916 static inline int thread_is_init(struct ptlrpc_thread *thread)
1918 return thread->t_flags == 0;
1921 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1923 return !!(thread->t_flags & SVC_STOPPED);
1926 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1928 return !!(thread->t_flags & SVC_STOPPING);
1931 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1933 return !!(thread->t_flags & SVC_STARTING);
1936 static inline int thread_is_running(struct ptlrpc_thread *thread)
1938 return !!(thread->t_flags & SVC_RUNNING);
1941 static inline int thread_is_event(struct ptlrpc_thread *thread)
1943 return !!(thread->t_flags & SVC_EVENT);
1946 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1948 return !!(thread->t_flags & SVC_SIGNAL);
1951 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1953 thread->t_flags &= ~flags;
1956 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1958 thread->t_flags = flags;
1961 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1963 thread->t_flags |= flags;
1966 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1969 if (thread->t_flags & flags) {
1970 thread->t_flags &= ~flags;
1977 * Request buffer descriptor structure.
1978 * This is a structure that contains one posted request buffer for service.
1979 * Once data land into a buffer, event callback creates actual request and
1980 * notifies wakes one of the service threads to process new incoming request.
1981 * More than one request can fit into the buffer.
1983 struct ptlrpc_request_buffer_desc {
1984 /** Link item for rqbds on a service */
1985 cfs_list_t rqbd_list;
1986 /** History of requests for this buffer */
1987 cfs_list_t rqbd_reqs;
1988 /** Back pointer to service for which this buffer is registered */
1989 struct ptlrpc_service_part *rqbd_svcpt;
1990 /** LNet descriptor */
1991 lnet_handle_md_t rqbd_md_h;
1993 /** The buffer itself */
1995 struct ptlrpc_cb_id rqbd_cbid;
1997 * This "embedded" request structure is only used for the
1998 * last request to fit into the buffer
2000 struct ptlrpc_request rqbd_req;
2003 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
2005 struct ptlrpc_service_ops {
2007 * if non-NULL called during thread creation (ptlrpc_start_thread())
2008 * to initialize service specific per-thread state.
2010 int (*so_thr_init)(struct ptlrpc_thread *thr);
2012 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2013 * destruct state created by ->srv_init().
2015 void (*so_thr_done)(struct ptlrpc_thread *thr);
2017 * Handler function for incoming requests for this service
2019 int (*so_req_handler)(struct ptlrpc_request *req);
2021 * function to determine priority of the request, it's called
2022 * on every new request
2024 int (*so_hpreq_handler)(struct ptlrpc_request *);
2026 * service-specific print fn
2028 void (*so_req_printer)(void *, struct ptlrpc_request *);
2031 #ifndef __cfs_cacheline_aligned
2032 /* NB: put it here for reducing patche dependence */
2033 # define __cfs_cacheline_aligned
2037 * How many high priority requests to serve before serving one normal
2040 #define PTLRPC_SVC_HP_RATIO 10
2043 * Definition of PortalRPC service.
2044 * The service is listening on a particular portal (like tcp port)
2045 * and perform actions for a specific server like IO service for OST
2046 * or general metadata service for MDS.
2048 struct ptlrpc_service {
2049 /** serialize /proc operations */
2050 spinlock_t srv_lock;
2051 /** most often accessed fields */
2052 /** chain thru all services */
2053 cfs_list_t srv_list;
2054 /** service operations table */
2055 struct ptlrpc_service_ops srv_ops;
2056 /** only statically allocated strings here; we don't clean them */
2058 /** only statically allocated strings here; we don't clean them */
2059 char *srv_thread_name;
2060 /** service thread list */
2061 cfs_list_t srv_threads;
2062 /** threads # should be created for each partition on initializing */
2063 int srv_nthrs_cpt_init;
2064 /** limit of threads number for each partition */
2065 int srv_nthrs_cpt_limit;
2066 /** Root of /proc dir tree for this service */
2067 cfs_proc_dir_entry_t *srv_procroot;
2068 /** Pointer to statistic data for this service */
2069 struct lprocfs_stats *srv_stats;
2070 /** # hp per lp reqs to handle */
2071 int srv_hpreq_ratio;
2072 /** biggest request to receive */
2073 int srv_max_req_size;
2074 /** biggest reply to send */
2075 int srv_max_reply_size;
2076 /** size of individual buffers */
2078 /** # buffers to allocate in 1 group */
2079 int srv_nbuf_per_group;
2080 /** Local portal on which to receive requests */
2081 __u32 srv_req_portal;
2082 /** Portal on the client to send replies to */
2083 __u32 srv_rep_portal;
2085 * Tags for lu_context associated with this thread, see struct
2089 /** soft watchdog timeout multiplier */
2090 int srv_watchdog_factor;
2091 /** under unregister_service */
2092 unsigned srv_is_stopping:1;
2094 /** max # request buffers in history per partition */
2095 int srv_hist_nrqbds_cpt_max;
2096 /** number of CPTs this service bound on */
2098 /** CPTs array this service bound on */
2100 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2102 /** CPT table this service is running over */
2103 struct cfs_cpt_table *srv_cptable;
2105 * partition data for ptlrpc service
2107 struct ptlrpc_service_part *srv_parts[0];
2111 * Definition of PortalRPC service partition data.
2112 * Although a service only has one instance of it right now, but we
2113 * will have multiple instances very soon (instance per CPT).
2115 * it has four locks:
2117 * serialize operations on rqbd and requests waiting for preprocess
2119 * serialize operations active requests sent to this portal
2121 * serialize adaptive timeout stuff
2123 * serialize operations on RS list (reply states)
2125 * We don't have any use-case to take two or more locks at the same time
2126 * for now, so there is no lock order issue.
2128 struct ptlrpc_service_part {
2129 /** back reference to owner */
2130 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2131 /* CPT id, reserved */
2133 /** always increasing number */
2135 /** # of starting threads */
2136 int scp_nthrs_starting;
2137 /** # of stopping threads, reserved for shrinking threads */
2138 int scp_nthrs_stopping;
2139 /** # running threads */
2140 int scp_nthrs_running;
2141 /** service threads list */
2142 cfs_list_t scp_threads;
2145 * serialize the following fields, used for protecting
2146 * rqbd list and incoming requests waiting for preprocess,
2147 * threads starting & stopping are also protected by this lock.
2149 spinlock_t scp_lock __cfs_cacheline_aligned;
2150 /** total # req buffer descs allocated */
2151 int scp_nrqbds_total;
2152 /** # posted request buffers for receiving */
2153 int scp_nrqbds_posted;
2154 /** in progress of allocating rqbd */
2155 int scp_rqbd_allocating;
2156 /** # incoming reqs */
2157 int scp_nreqs_incoming;
2158 /** request buffers to be reposted */
2159 cfs_list_t scp_rqbd_idle;
2160 /** req buffers receiving */
2161 cfs_list_t scp_rqbd_posted;
2162 /** incoming reqs */
2163 cfs_list_t scp_req_incoming;
2164 /** timeout before re-posting reqs, in tick */
2165 cfs_duration_t scp_rqbd_timeout;
2167 * all threads sleep on this. This wait-queue is signalled when new
2168 * incoming request arrives and when difficult reply has to be handled.
2170 cfs_waitq_t scp_waitq;
2172 /** request history */
2173 cfs_list_t scp_hist_reqs;
2174 /** request buffer history */
2175 cfs_list_t scp_hist_rqbds;
2176 /** # request buffers in history */
2177 int scp_hist_nrqbds;
2178 /** sequence number for request */
2180 /** highest seq culled from history */
2181 __u64 scp_hist_seq_culled;
2184 * serialize the following fields, used for processing requests
2185 * sent to this portal
2187 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2188 /** # reqs in either of the NRS heads below */
2189 /** # reqs being served */
2190 int scp_nreqs_active;
2191 /** # HPreqs being served */
2192 int scp_nhreqs_active;
2193 /** # hp requests handled */
2196 /** NRS head for regular requests */
2197 struct ptlrpc_nrs scp_nrs_reg;
2198 /** NRS head for HP requests; this is only valid for services that can
2199 * handle HP requests */
2200 struct ptlrpc_nrs *scp_nrs_hp;
2205 * serialize the following fields, used for changes on
2208 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2209 /** estimated rpc service time */
2210 struct adaptive_timeout scp_at_estimate;
2211 /** reqs waiting for replies */
2212 struct ptlrpc_at_array scp_at_array;
2213 /** early reply timer */
2214 cfs_timer_t scp_at_timer;
2216 cfs_time_t scp_at_checktime;
2217 /** check early replies */
2218 unsigned scp_at_check;
2222 * serialize the following fields, used for processing
2223 * replies for this portal
2225 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2226 /** all the active replies */
2227 cfs_list_t scp_rep_active;
2229 /** replies waiting for service */
2230 cfs_list_t scp_rep_queue;
2232 /** List of free reply_states */
2233 cfs_list_t scp_rep_idle;
2234 /** waitq to run, when adding stuff to srv_free_rs_list */
2235 cfs_waitq_t scp_rep_waitq;
2236 /** # 'difficult' replies */
2237 cfs_atomic_t scp_nreps_difficult;
2240 #define ptlrpc_service_for_each_part(part, i, svc) \
2242 i < (svc)->srv_ncpts && \
2243 (svc)->srv_parts != NULL && \
2244 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2247 * Declaration of ptlrpcd control structure
2249 struct ptlrpcd_ctl {
2251 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2253 unsigned long pc_flags;
2255 * Thread lock protecting structure fields.
2261 struct completion pc_starting;
2265 struct completion pc_finishing;
2267 * Thread requests set.
2269 struct ptlrpc_request_set *pc_set;
2271 * Thread name used in cfs_daemonize()
2275 * Environment for request interpreters to run in.
2277 struct lu_env pc_env;
2279 * Index of ptlrpcd thread in the array.
2283 * Number of the ptlrpcd's partners.
2287 * Pointer to the array of partners' ptlrpcd_ctl structure.
2289 struct ptlrpcd_ctl **pc_partners;
2291 * Record the partner index to be processed next.
2296 * Async rpcs flag to make sure that ptlrpcd_check() is called only
2301 * Currently not used.
2305 * User-space async rpcs callback.
2307 void *pc_wait_callback;
2309 * User-space check idle rpcs callback.
2311 void *pc_idle_callback;
2315 /* Bits for pc_flags */
2316 enum ptlrpcd_ctl_flags {
2318 * Ptlrpc thread start flag.
2320 LIOD_START = 1 << 0,
2322 * Ptlrpc thread stop flag.
2326 * Ptlrpc thread force flag (only stop force so far).
2327 * This will cause aborting any inflight rpcs handled
2328 * by thread if LIOD_STOP is specified.
2330 LIOD_FORCE = 1 << 2,
2332 * This is a recovery ptlrpc thread.
2334 LIOD_RECOVERY = 1 << 3,
2336 * The ptlrpcd is bound to some CPU core.
2345 * Service compatibility function; policy is compatible with all services.
2347 * \param[in] svc The service the policy is attempting to register with.
2348 * \param[in] desc The policy descriptor
2350 * \retval true The policy is compatible with the NRS head
2352 * \see ptlrpc_nrs_pol_desc::pd_compat()
2355 nrs_policy_compat_all(struct ptlrpc_service *svc,
2356 const struct ptlrpc_nrs_pol_desc *desc)
2362 * Service compatibility function; policy is compatible with only a specific
2363 * service which is identified by its human-readable name at
2364 * ptlrpc_service::srv_name.
2366 * \param[in] svc The service the policy is attempting to register with.
2367 * \param[in] desc The policy descriptor
2369 * \retval false The policy is not compatible with the NRS head
2370 * \retval true The policy is compatible with the NRS head
2372 * \see ptlrpc_nrs_pol_desc::pd_compat()
2375 nrs_policy_compat_one(struct ptlrpc_service *svc,
2376 const struct ptlrpc_nrs_pol_desc *desc)
2378 LASSERT(desc->pd_compat_svc_name != NULL);
2379 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2384 /* ptlrpc/events.c */
2385 extern lnet_handle_eq_t ptlrpc_eq_h;
2386 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2387 lnet_process_id_t *peer, lnet_nid_t *self);
2389 * These callbacks are invoked by LNet when something happened to
2393 extern void request_out_callback(lnet_event_t *ev);
2394 extern void reply_in_callback(lnet_event_t *ev);
2395 extern void client_bulk_callback(lnet_event_t *ev);
2396 extern void request_in_callback(lnet_event_t *ev);
2397 extern void reply_out_callback(lnet_event_t *ev);
2398 #ifdef HAVE_SERVER_SUPPORT
2399 extern void server_bulk_callback(lnet_event_t *ev);
2403 /* ptlrpc/connection.c */
2404 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2406 struct obd_uuid *uuid);
2407 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2408 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2409 int ptlrpc_connection_init(void);
2410 void ptlrpc_connection_fini(void);
2411 extern lnet_pid_t ptl_get_pid(void);
2413 /* ptlrpc/niobuf.c */
2415 * Actual interfacing with LNet to put/get/register/unregister stuff
2418 #ifdef HAVE_SERVER_SUPPORT
2419 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
2420 unsigned npages, unsigned max_brw,
2421 unsigned type, unsigned portal);
2422 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc);
2423 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc);
2425 static inline int ptlrpc_server_bulk_active(struct ptlrpc_bulk_desc *desc)
2429 LASSERT(desc != NULL);
2431 spin_lock(&desc->bd_lock);
2432 rc = desc->bd_md_count;
2433 spin_unlock(&desc->bd_lock);
2438 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2439 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2441 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2443 struct ptlrpc_bulk_desc *desc;
2446 LASSERT(req != NULL);
2447 desc = req->rq_bulk;
2449 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2450 req->rq_bulk_deadline > cfs_time_current_sec())
2456 spin_lock(&desc->bd_lock);
2457 rc = desc->bd_md_count;
2458 spin_unlock(&desc->bd_lock);
2462 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2463 #define PTLRPC_REPLY_EARLY 0x02
2464 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2465 int ptlrpc_reply(struct ptlrpc_request *req);
2466 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2467 int ptlrpc_error(struct ptlrpc_request *req);
2468 void ptlrpc_resend_req(struct ptlrpc_request *request);
2469 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2470 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2471 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2474 /* ptlrpc/client.c */
2476 * Client-side portals API. Everything to send requests, receive replies,
2477 * request queues, request management, etc.
2480 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2481 struct ptlrpc_client *);
2482 void ptlrpc_cleanup_client(struct obd_import *imp);
2483 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2485 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2486 int ptlrpc_replay_req(struct ptlrpc_request *req);
2487 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2488 void ptlrpc_restart_req(struct ptlrpc_request *req);
2489 void ptlrpc_abort_inflight(struct obd_import *imp);
2490 void ptlrpc_cleanup_imp(struct obd_import *imp);
2491 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2493 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2494 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2496 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2497 set_interpreter_func fn, void *data);
2498 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2499 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2500 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2501 int ptlrpc_expired_set(void *data);
2502 void ptlrpc_interrupted_set(void *data);
2503 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2504 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2505 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2506 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2507 struct ptlrpc_request *req);
2509 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2510 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2512 struct ptlrpc_request_pool *
2513 ptlrpc_init_rq_pool(int, int,
2514 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2516 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2517 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2518 const struct req_format *format);
2519 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2520 struct ptlrpc_request_pool *,
2521 const struct req_format *format);
2522 void ptlrpc_request_free(struct ptlrpc_request *request);
2523 int ptlrpc_request_pack(struct ptlrpc_request *request,
2524 __u32 version, int opcode);
2525 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2526 const struct req_format *format,
2527 __u32 version, int opcode);
2528 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2529 __u32 version, int opcode, char **bufs,
2530 struct ptlrpc_cli_ctx *ctx);
2531 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2532 int opcode, int count, __u32 *lengths,
2534 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2535 __u32 version, int opcode,
2536 int count, __u32 *lengths, char **bufs,
2537 struct ptlrpc_request_pool *pool);
2538 void ptlrpc_req_finished(struct ptlrpc_request *request);
2539 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2540 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2541 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2542 unsigned npages, unsigned max_brw,
2543 unsigned type, unsigned portal);
2544 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2545 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2547 __ptlrpc_free_bulk(bulk, 1);
2549 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2551 __ptlrpc_free_bulk(bulk, 0);
2553 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2554 cfs_page_t *page, int pageoffset, int len, int);
2555 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2556 cfs_page_t *page, int pageoffset,
2559 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2562 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2563 cfs_page_t *page, int pageoffset,
2566 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2569 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2570 struct obd_import *imp);
2571 __u64 ptlrpc_next_xid(void);
2572 __u64 ptlrpc_sample_next_xid(void);
2573 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2575 /* Set of routines to run a function in ptlrpcd context */
2576 void *ptlrpcd_alloc_work(struct obd_import *imp,
2577 int (*cb)(const struct lu_env *, void *), void *data);
2578 void ptlrpcd_destroy_work(void *handler);
2579 int ptlrpcd_queue_work(void *handler);
2582 struct ptlrpc_service_buf_conf {
2583 /* nbufs is how many buffers to post */
2584 unsigned int bc_nbufs;
2585 /* buffer size to post */
2586 unsigned int bc_buf_size;
2587 /* portal to listed for requests on */
2588 unsigned int bc_req_portal;
2589 /* portal of where to send replies to */
2590 unsigned int bc_rep_portal;
2591 /* maximum request size to be accepted for this service */
2592 unsigned int bc_req_max_size;
2593 /* maximum reply size this service can ever send */
2594 unsigned int bc_rep_max_size;
2597 struct ptlrpc_service_thr_conf {
2598 /* threadname should be 8 characters or less - 6 will be added on */
2600 /* threads increasing factor for each CPU */
2601 unsigned int tc_thr_factor;
2602 /* service threads # to start on each partition while initializing */
2603 unsigned int tc_nthrs_init;
2605 * low water of threads # upper-limit on each partition while running,
2606 * service availability may be impacted if threads number is lower
2607 * than this value. It can be ZERO if the service doesn't require
2608 * CPU affinity or there is only one partition.
2610 unsigned int tc_nthrs_base;
2611 /* "soft" limit for total threads number */
2612 unsigned int tc_nthrs_max;
2613 /* user specified threads number, it will be validated due to
2614 * other members of this structure. */
2615 unsigned int tc_nthrs_user;
2616 /* set NUMA node affinity for service threads */
2617 unsigned int tc_cpu_affinity;
2618 /* Tags for lu_context associated with service thread */
2622 struct ptlrpc_service_cpt_conf {
2623 struct cfs_cpt_table *cc_cptable;
2624 /* string pattern to describe CPTs for a service */
2628 struct ptlrpc_service_conf {
2631 /* soft watchdog timeout multiplifier to print stuck service traces */
2632 unsigned int psc_watchdog_factor;
2633 /* buffer information */
2634 struct ptlrpc_service_buf_conf psc_buf;
2635 /* thread information */
2636 struct ptlrpc_service_thr_conf psc_thr;
2637 /* CPU partition information */
2638 struct ptlrpc_service_cpt_conf psc_cpt;
2639 /* function table */
2640 struct ptlrpc_service_ops psc_ops;
2643 /* ptlrpc/service.c */
2645 * Server-side services API. Register/unregister service, request state
2646 * management, service thread management
2650 void ptlrpc_save_lock(struct ptlrpc_request *req,
2651 struct lustre_handle *lock, int mode, int no_ack);
2652 void ptlrpc_commit_replies(struct obd_export *exp);
2653 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2654 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2655 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2656 struct ptlrpc_service *ptlrpc_register_service(
2657 struct ptlrpc_service_conf *conf,
2658 struct proc_dir_entry *proc_entry);
2659 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2661 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2662 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2663 int liblustre_check_services(void *arg);
2664 void ptlrpc_daemonize(char *name);
2665 int ptlrpc_service_health_check(struct ptlrpc_service *);
2666 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2669 int ptlrpc_hr_init(void);
2670 void ptlrpc_hr_fini(void);
2672 # define ptlrpc_hr_init() (0)
2673 # define ptlrpc_hr_fini() do {} while(0)
2678 /* ptlrpc/import.c */
2683 int ptlrpc_connect_import(struct obd_import *imp);
2684 int ptlrpc_init_import(struct obd_import *imp);
2685 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2686 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2687 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2690 /* ptlrpc/pack_generic.c */
2691 int ptlrpc_reconnect_import(struct obd_import *imp);
2695 * ptlrpc msg buffer and swab interface
2699 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2701 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2703 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2704 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2706 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2707 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2709 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2710 __u32 *lens, char **bufs);
2711 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2713 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2714 __u32 *lens, char **bufs, int flags);
2715 #define LPRFL_EARLY_REPLY 1
2716 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2717 char **bufs, int flags);
2718 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2719 unsigned int newlen, int move_data);
2720 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2721 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2722 int lustre_msg_hdr_size(__u32 magic, int count);
2723 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2724 int lustre_msg_size_v2(int count, __u32 *lengths);
2725 int lustre_packed_msg_size(struct lustre_msg *msg);
2726 int lustre_msg_early_size(void);
2727 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2728 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2729 int lustre_msg_buflen(struct lustre_msg *m, int n);
2730 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2731 int lustre_msg_bufcount(struct lustre_msg *m);
2732 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2733 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2734 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2735 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2736 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2737 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2738 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2739 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2740 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2741 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2742 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2743 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2744 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2745 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2746 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2747 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2748 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2749 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2750 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2751 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2752 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2753 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2754 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2755 int lustre_msg_get_status(struct lustre_msg *msg);
2756 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2757 int lustre_msg_is_v1(struct lustre_msg *msg);
2758 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2759 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2760 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2761 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2762 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2763 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2764 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18);
2766 # warning "remove checksum compatibility support for b1_8"
2767 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2769 void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle);
2770 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2771 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2772 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2773 void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed);
2774 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2775 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2776 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2777 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2778 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2779 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2780 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2781 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2782 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2783 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2786 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2787 unsigned int newlen, int move_data)
2789 LASSERT(req->rq_reply_state);
2790 LASSERT(req->rq_repmsg);
2791 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2796 /** Change request phase of \a req to \a new_phase */
2798 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2800 if (req->rq_phase == new_phase)
2803 if (new_phase == RQ_PHASE_UNREGISTERING) {
2804 req->rq_next_phase = req->rq_phase;
2806 cfs_atomic_inc(&req->rq_import->imp_unregistering);
2809 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2811 cfs_atomic_dec(&req->rq_import->imp_unregistering);
2814 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2815 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2817 req->rq_phase = new_phase;
2821 * Returns true if request \a req got early reply and hard deadline is not met
2824 ptlrpc_client_early(struct ptlrpc_request *req)
2826 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2827 req->rq_reply_deadline > cfs_time_current_sec())
2829 return req->rq_early;
2833 * Returns true if we got real reply from server for this request
2836 ptlrpc_client_replied(struct ptlrpc_request *req)
2838 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2839 req->rq_reply_deadline > cfs_time_current_sec())
2841 return req->rq_replied;
2844 /** Returns true if request \a req is in process of receiving server reply */
2846 ptlrpc_client_recv(struct ptlrpc_request *req)
2848 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2849 req->rq_reply_deadline > cfs_time_current_sec())
2851 return req->rq_receiving_reply;
2855 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2859 spin_lock(&req->rq_lock);
2860 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2861 req->rq_reply_deadline > cfs_time_current_sec()) {
2862 spin_unlock(&req->rq_lock);
2865 rc = req->rq_receiving_reply || req->rq_must_unlink;
2866 spin_unlock(&req->rq_lock);
2871 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2873 if (req->rq_set == NULL)
2874 cfs_waitq_signal(&req->rq_reply_waitq);
2876 cfs_waitq_signal(&req->rq_set->set_waitq);
2880 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2882 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2883 cfs_atomic_inc(&rs->rs_refcount);
2887 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2889 LASSERT(cfs_atomic_read(&rs->rs_refcount) > 0);
2890 if (cfs_atomic_dec_and_test(&rs->rs_refcount))
2891 lustre_free_reply_state(rs);
2894 /* Should only be called once per req */
2895 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2897 if (req->rq_reply_state == NULL)
2898 return; /* shouldn't occur */
2899 ptlrpc_rs_decref(req->rq_reply_state);
2900 req->rq_reply_state = NULL;
2901 req->rq_repmsg = NULL;
2904 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2906 return lustre_msg_get_magic(req->rq_reqmsg);
2909 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2911 switch (req->rq_reqmsg->lm_magic) {
2912 case LUSTRE_MSG_MAGIC_V2:
2913 return req->rq_reqmsg->lm_repsize;
2915 LASSERTF(0, "incorrect message magic: %08x\n",
2916 req->rq_reqmsg->lm_magic);
2921 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2923 if (req->rq_delay_limit != 0 &&
2924 cfs_time_before(cfs_time_add(req->rq_queued_time,
2925 cfs_time_seconds(req->rq_delay_limit)),
2926 cfs_time_current())) {
2932 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2934 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2935 spin_lock(&req->rq_lock);
2936 req->rq_no_resend = 1;
2937 spin_unlock(&req->rq_lock);
2939 return req->rq_no_resend;
2943 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2945 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2947 return svcpt->scp_service->srv_watchdog_factor *
2948 max_t(int, at, obd_timeout);
2951 static inline struct ptlrpc_service *
2952 ptlrpc_req2svc(struct ptlrpc_request *req)
2954 LASSERT(req->rq_rqbd != NULL);
2955 return req->rq_rqbd->rqbd_svcpt->scp_service;
2958 /* ldlm/ldlm_lib.c */
2960 * Target client logic
2963 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2964 int client_obd_cleanup(struct obd_device *obddev);
2965 int client_connect_import(const struct lu_env *env,
2966 struct obd_export **exp, struct obd_device *obd,
2967 struct obd_uuid *cluuid, struct obd_connect_data *,
2969 int client_disconnect_export(struct obd_export *exp);
2970 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2972 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2973 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2974 struct obd_uuid *uuid);
2975 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2976 void client_destroy_import(struct obd_import *imp);
2979 #ifdef HAVE_SERVER_SUPPORT
2980 int server_disconnect_export(struct obd_export *exp);
2983 /* ptlrpc/pinger.c */
2985 * Pinger API (client side only)
2988 extern int suppress_pings;
2989 enum timeout_event {
2992 struct timeout_item;
2993 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2994 int ptlrpc_pinger_add_import(struct obd_import *imp);
2995 int ptlrpc_pinger_del_import(struct obd_import *imp);
2996 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
2997 timeout_cb_t cb, void *data,
2998 cfs_list_t *obd_list);
2999 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
3000 enum timeout_event event);
3001 struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp);
3002 int ptlrpc_obd_ping(struct obd_device *obd);
3003 cfs_time_t ptlrpc_suspend_wakeup_time(void);
3005 void ping_evictor_start(void);
3006 void ping_evictor_stop(void);
3008 #define ping_evictor_start() do {} while (0)
3009 #define ping_evictor_stop() do {} while (0)
3011 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req);
3014 /* ptlrpc daemon bind policy */
3016 /* all ptlrpcd threads are free mode */
3017 PDB_POLICY_NONE = 1,
3018 /* all ptlrpcd threads are bound mode */
3019 PDB_POLICY_FULL = 2,
3020 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3021 PDB_POLICY_PAIR = 3,
3022 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3023 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3024 * If kernel supports NUMA, pthrpcd threads are binded and
3025 * grouped by NUMA node */
3026 PDB_POLICY_NEIGHBOR = 4,
3029 /* ptlrpc daemon load policy
3030 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3031 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3032 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3033 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3034 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3036 /* on the same CPU core as the caller */
3037 PDL_POLICY_SAME = 1,
3038 /* within the same CPU partition, but not the same core as the caller */
3039 PDL_POLICY_LOCAL = 2,
3040 /* round-robin on all CPU cores, but not the same core as the caller */
3041 PDL_POLICY_ROUND = 3,
3042 /* the specified CPU core is preferred, but not enforced */
3043 PDL_POLICY_PREFERRED = 4,
3046 /* ptlrpc/ptlrpcd.c */
3047 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
3048 void ptlrpcd_wake(struct ptlrpc_request *req);
3049 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
3050 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
3051 int ptlrpcd_addref(void);
3052 void ptlrpcd_decref(void);
3054 /* ptlrpc/lproc_ptlrpc.c */
3056 * procfs output related functions
3059 const char* ll_opcode2str(__u32 opcode);
3061 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
3062 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
3063 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
3065 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
3066 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
3067 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
3071 /* ptlrpc/llog_server.c */
3072 int llog_origin_handle_open(struct ptlrpc_request *req);
3073 int llog_origin_handle_destroy(struct ptlrpc_request *req);
3074 int llog_origin_handle_prev_block(struct ptlrpc_request *req);
3075 int llog_origin_handle_next_block(struct ptlrpc_request *req);
3076 int llog_origin_handle_read_header(struct ptlrpc_request *req);
3077 int llog_origin_handle_close(struct ptlrpc_request *req);
3078 int llog_origin_handle_cancel(struct ptlrpc_request *req);
3080 /* ptlrpc/llog_client.c */
3081 extern struct llog_operations llog_client_ops;