4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #include "../../include/linux/libcfs/libcfs.h"
60 #include "../../include/linux/lnet/lnet.h"
61 #include "lustre/lustre_idl.h"
62 #include "lustre_ha.h"
63 #include "lustre_sec.h"
64 #include "lustre_import.h"
65 #include "lprocfs_status.h"
66 #include "lu_object.h"
67 #include "lustre_req_layout.h"
69 #include "obd_support.h"
70 #include "lustre_ver.h"
72 /* MD flags we _always_ use */
73 #define PTLRPC_MD_OPTIONS 0
76 * Max # of bulk operations in one request.
77 * In order for the client and server to properly negotiate the maximum
78 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
79 * value. The client is free to limit the actual RPC size for any bulk
80 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
81 #define PTLRPC_BULK_OPS_BITS 2
82 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
84 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
85 * should not be used on the server at all. Otherwise, it imposes a
86 * protocol limitation on the maximum RPC size that can be used by any
87 * RPC sent to that server in the future. Instead, the server should
88 * use the negotiated per-client ocd_brw_size to determine the bulk
90 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
93 * Define maxima for bulk I/O.
95 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
96 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
97 * currently supported maximum between peers at connect via ocd_brw_size.
99 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
100 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
101 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
103 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
104 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
105 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
106 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
107 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
108 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
110 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
111 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
112 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
114 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
115 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
117 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
118 # error "PTLRPC_MAX_BRW_SIZE too big"
120 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
121 # error "PTLRPC_MAX_BRW_PAGES too big"
124 #define PTLRPC_NTHRS_INIT 2
129 * Constants determine how memory is used to buffer incoming service requests.
131 * ?_NBUFS # buffers to allocate when growing the pool
132 * ?_BUFSIZE # bytes in a single request buffer
133 * ?_MAXREQSIZE # maximum request service will receive
135 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
136 * of ?_NBUFS is added to the pool.
138 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
139 * considered full when less than ?_MAXREQSIZE is left in them.
144 * Constants determine how threads are created for ptlrpc service.
146 * ?_NTHRS_INIT # threads to create for each service partition on
147 * initializing. If it's non-affinity service and
148 * there is only one partition, it's the overall #
149 * threads for the service while initializing.
150 * ?_NTHRS_BASE # threads should be created at least for each
151 * ptlrpc partition to keep the service healthy.
152 * It's the low-water mark of threads upper-limit
153 * for each partition.
154 * ?_THR_FACTOR # threads can be added on threads upper-limit for
155 * each CPU core. This factor is only for reference,
156 * we might decrease value of factor if number of cores
157 * per CPT is above a limit.
158 * ?_NTHRS_MAX # overall threads can be created for a service,
159 * it's a soft limit because if service is running
160 * on machine with hundreds of cores and tens of
161 * CPU partitions, we need to guarantee each partition
162 * has ?_NTHRS_BASE threads, which means total threads
163 * will be ?_NTHRS_BASE * number_of_cpts which can
164 * exceed ?_NTHRS_MAX.
168 * #define MDS_NTHRS_INIT 2
169 * #define MDS_NTHRS_BASE 64
170 * #define MDS_NTHRS_FACTOR 8
171 * #define MDS_NTHRS_MAX 1024
174 * ---------------------------------------------------------------------
175 * Server(A) has 16 cores, user configured it to 4 partitions so each
176 * partition has 4 cores, then actual number of service threads on each
178 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
180 * Total number of threads for the service is:
181 * 96 * partitions(4) = 384
184 * ---------------------------------------------------------------------
185 * Server(B) has 32 cores, user configured it to 4 partitions so each
186 * partition has 8 cores, then actual number of service threads on each
188 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
190 * Total number of threads for the service is:
191 * 128 * partitions(4) = 512
194 * ---------------------------------------------------------------------
195 * Server(B) has 96 cores, user configured it to 8 partitions so each
196 * partition has 12 cores, then actual number of service threads on each
198 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
200 * Total number of threads for the service is:
201 * 160 * partitions(8) = 1280
203 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
204 * as upper limit of threads number for each partition:
205 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
208 * ---------------------------------------------------------------------
209 * Server(C) have a thousand of cores and user configured it to 32 partitions
210 * MDS_NTHRS_BASE(64) * 32 = 2048
212 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
213 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
214 * to keep service healthy, so total number of threads will just be 2048.
216 * NB: we don't suggest to choose server with that many cores because backend
217 * filesystem itself, buffer cache, or underlying network stack might
218 * have some SMP scalability issues at that large scale.
220 * If user already has a fat machine with hundreds or thousands of cores,
221 * there are two choices for configuration:
222 * a) create CPU table from subset of all CPUs and run Lustre on
224 * b) bind service threads on a few partitions, see modparameters of
225 * MDS and OSS for details
227 * NB: these calculations (and examples below) are simplified to help
228 * understanding, the real implementation is a little more complex,
229 * please see ptlrpc_server_nthreads_check() for details.
234 * LDLM threads constants:
236 * Given 8 as factor and 24 as base threads number
239 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
242 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
243 * threads for each partition and total threads number will be 112.
246 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
247 * threads for each partition to keep service healthy, so total threads
248 * number should be 24 * 8 = 192.
250 * So with these constants, threads number will be at the similar level
251 * of old versions, unless target machine has over a hundred cores
253 #define LDLM_THR_FACTOR 8
254 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
255 #define LDLM_NTHRS_BASE 24
256 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
258 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
259 #define LDLM_CLIENT_NBUFS 1
260 #define LDLM_SERVER_NBUFS 64
261 #define LDLM_BUFSIZE (8 * 1024)
262 #define LDLM_MAXREQSIZE (5 * 1024)
263 #define LDLM_MAXREPSIZE (1024)
265 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
267 #define OST_MAXREQSIZE (5 * 1024)
269 /* Macro to hide a typecast. */
270 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
273 * Structure to single define portal connection.
275 struct ptlrpc_connection {
276 /** linkage for connections hash table */
277 struct hlist_node c_hash;
278 /** Our own lnet nid for this connection */
280 /** Remote side nid for this connection */
281 lnet_process_id_t c_peer;
282 /** UUID of the other side */
283 struct obd_uuid c_remote_uuid;
284 /** reference counter for this connection */
288 /** Client definition for PortalRPC */
289 struct ptlrpc_client {
290 /** What lnet portal does this client send messages to by default */
291 __u32 cli_request_portal;
292 /** What portal do we expect replies on */
293 __u32 cli_reply_portal;
294 /** Name of the client */
298 /** state flags of requests */
299 /* XXX only ones left are those used by the bulk descs as well! */
300 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
301 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
303 #define REQ_MAX_ACK_LOCKS 8
305 union ptlrpc_async_args {
307 * Scratchpad for passing args to completion interpreter. Users
308 * cast to the struct of their choosing, and CLASSERT that this is
309 * big enough. For _tons_ of context, OBD_ALLOC a struct and store
310 * a pointer to it here. The pointer_arg ensures this struct is at
311 * least big enough for that.
313 void *pointer_arg[11];
317 struct ptlrpc_request_set;
318 typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int);
319 typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *);
322 * Definition of request set structure.
323 * Request set is a list of requests (not necessary to the same target) that
324 * once populated with RPCs could be sent in parallel.
325 * There are two kinds of request sets. General purpose and with dedicated
326 * serving thread. Example of the latter is ptlrpcd set.
327 * For general purpose sets once request set started sending it is impossible
328 * to add new requests to such set.
329 * Provides a way to call "completion callbacks" when all requests in the set
332 struct ptlrpc_request_set {
333 atomic_t set_refcount;
334 /** number of in queue requests */
335 atomic_t set_new_count;
336 /** number of uncompleted requests */
337 atomic_t set_remaining;
338 /** wait queue to wait on for request events */
339 wait_queue_head_t set_waitq;
340 wait_queue_head_t *set_wakeup_ptr;
341 /** List of requests in the set */
342 struct list_head set_requests;
344 * List of completion callbacks to be called when the set is completed
345 * This is only used if \a set_interpret is NULL.
346 * Links struct ptlrpc_set_cbdata.
348 struct list_head set_cblist;
349 /** Completion callback, if only one. */
350 set_interpreter_func set_interpret;
351 /** opaq argument passed to completion \a set_interpret callback. */
354 * Lock for \a set_new_requests manipulations
355 * locked so that any old caller can communicate requests to
356 * the set holder who can then fold them into the lock-free set
358 spinlock_t set_new_req_lock;
359 /** List of new yet unsent requests. Only used with ptlrpcd now. */
360 struct list_head set_new_requests;
362 /** rq_status of requests that have been freed already */
364 /** Additional fields used by the flow control extension */
365 /** Maximum number of RPCs in flight */
366 int set_max_inflight;
367 /** Callback function used to generate RPCs */
368 set_producer_func set_producer;
369 /** opaq argument passed to the producer callback */
370 void *set_producer_arg;
374 * Description of a single ptrlrpc_set callback
376 struct ptlrpc_set_cbdata {
377 /** List linkage item */
378 struct list_head psc_item;
379 /** Pointer to interpreting function */
380 set_interpreter_func psc_interpret;
381 /** Opaq argument to pass to the callback */
385 struct ptlrpc_bulk_desc;
386 struct ptlrpc_service_part;
387 struct ptlrpc_service;
390 * ptlrpc callback & work item stuff
392 struct ptlrpc_cb_id {
393 void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */
394 void *cbid_arg; /* additional arg */
397 /** Maximum number of locks to fit into reply state */
398 #define RS_MAX_LOCKS 8
402 * Structure to define reply state on the server
403 * Reply state holds various reply message information. Also for "difficult"
404 * replies (rep-ack case) we store the state after sending reply and wait
405 * for the client to acknowledge the reception. In these cases locks could be
406 * added to the state for replay/failover consistency guarantees.
408 struct ptlrpc_reply_state {
409 /** Callback description */
410 struct ptlrpc_cb_id rs_cb_id;
411 /** Linkage for list of all reply states in a system */
412 struct list_head rs_list;
413 /** Linkage for list of all reply states on same export */
414 struct list_head rs_exp_list;
415 /** Linkage for list of all reply states for same obd */
416 struct list_head rs_obd_list;
418 struct list_head rs_debug_list;
420 /** A spinlock to protect the reply state flags */
422 /** Reply state flags */
423 unsigned long rs_difficult:1; /* ACK/commit stuff */
424 unsigned long rs_no_ack:1; /* no ACK, even for
425 difficult requests */
426 unsigned long rs_scheduled:1; /* being handled? */
427 unsigned long rs_scheduled_ever:1;/* any schedule attempts? */
428 unsigned long rs_handled:1; /* been handled yet? */
429 unsigned long rs_on_net:1; /* reply_out_callback pending? */
430 unsigned long rs_prealloc:1; /* rs from prealloc list */
431 unsigned long rs_committed:1;/* the transaction was committed
432 and the rs was dispatched
433 by ptlrpc_commit_replies */
434 /** Size of the state */
438 /** Transaction number */
442 struct obd_export *rs_export;
443 struct ptlrpc_service_part *rs_svcpt;
444 /** Lnet metadata handle for the reply */
445 lnet_handle_md_t rs_md_h;
446 atomic_t rs_refcount;
448 /** Context for the service thread */
449 struct ptlrpc_svc_ctx *rs_svc_ctx;
450 /** Reply buffer (actually sent to the client), encoded if needed */
451 struct lustre_msg *rs_repbuf; /* wrapper */
452 /** Size of the reply buffer */
453 int rs_repbuf_len; /* wrapper buf length */
454 /** Size of the reply message */
455 int rs_repdata_len; /* wrapper msg length */
457 * Actual reply message. Its content is encrypted (if needed) to
458 * produce reply buffer for actual sending. In simple case
459 * of no network encryption we just set \a rs_repbuf to \a rs_msg
461 struct lustre_msg *rs_msg; /* reply message */
463 /** Number of locks awaiting client ACK */
465 /** Handles of locks awaiting client reply ACK */
466 struct lustre_handle rs_locks[RS_MAX_LOCKS];
467 /** Lock modes of locks in \a rs_locks */
468 ldlm_mode_t rs_modes[RS_MAX_LOCKS];
471 struct ptlrpc_thread;
475 RQ_PHASE_NEW = 0xebc0de00,
476 RQ_PHASE_RPC = 0xebc0de01,
477 RQ_PHASE_BULK = 0xebc0de02,
478 RQ_PHASE_INTERPRET = 0xebc0de03,
479 RQ_PHASE_COMPLETE = 0xebc0de04,
480 RQ_PHASE_UNREGISTERING = 0xebc0de05,
481 RQ_PHASE_UNDEFINED = 0xebc0de06
484 /** Type of request interpreter call-back */
485 typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env,
486 struct ptlrpc_request *req,
490 * Definition of request pool structure.
491 * The pool is used to store empty preallocated requests for the case
492 * when we would actually need to send something without performing
493 * any allocations (to avoid e.g. OOM).
495 struct ptlrpc_request_pool {
496 /** Locks the list */
498 /** list of ptlrpc_request structs */
499 struct list_head prp_req_list;
500 /** Maximum message size that would fit into a request from this pool */
502 /** Function to allocate more requests for this pool */
503 void (*prp_populate)(struct ptlrpc_request_pool *, int);
512 * \defgroup nrs Network Request Scheduler
515 struct ptlrpc_nrs_policy;
516 struct ptlrpc_nrs_resource;
517 struct ptlrpc_nrs_request;
520 * NRS control operations.
522 * These are common for all policies.
524 enum ptlrpc_nrs_ctl {
526 * Not a valid opcode.
528 PTLRPC_NRS_CTL_INVALID,
530 * Activate the policy.
532 PTLRPC_NRS_CTL_START,
534 * Reserved for multiple primary policies, which may be a possibility
539 * Policies can start using opcodes from this value and onwards for
540 * their own purposes; the assigned value itself is arbitrary.
542 PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20,
546 * ORR policy operations
549 NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC,
550 NRS_CTL_ORR_WR_QUANTUM,
551 NRS_CTL_ORR_RD_OFF_TYPE,
552 NRS_CTL_ORR_WR_OFF_TYPE,
553 NRS_CTL_ORR_RD_SUPP_REQ,
554 NRS_CTL_ORR_WR_SUPP_REQ,
558 * NRS policy operations.
560 * These determine the behaviour of a policy, and are called in response to
563 struct ptlrpc_nrs_pol_ops {
565 * Called during policy registration; this operation is optional.
567 * \param[in,out] policy The policy being initialized
569 int (*op_policy_init) (struct ptlrpc_nrs_policy *policy);
571 * Called during policy unregistration; this operation is optional.
573 * \param[in,out] policy The policy being unregistered/finalized
575 void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy);
577 * Called when activating a policy via lprocfs; policies allocate and
578 * initialize their resources here; this operation is optional.
580 * \param[in,out] policy The policy being started
582 * \see nrs_policy_start_locked()
584 int (*op_policy_start) (struct ptlrpc_nrs_policy *policy);
586 * Called when deactivating a policy via lprocfs; policies deallocate
587 * their resources here; this operation is optional
589 * \param[in,out] policy The policy being stopped
591 * \see nrs_policy_stop0()
593 void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy);
595 * Used for policy-specific operations; i.e. not generic ones like
596 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
597 * to an ioctl; this operation is optional.
599 * \param[in,out] policy The policy carrying out operation \a opc
600 * \param[in] opc The command operation being carried out
601 * \param[in,out] arg An generic buffer for communication between the
602 * user and the control operation
607 * \see ptlrpc_nrs_policy_control()
609 int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy,
610 enum ptlrpc_nrs_ctl opc, void *arg);
613 * Called when obtaining references to the resources of the resource
614 * hierarchy for a request that has arrived for handling at the PTLRPC
615 * service. Policies should return -ve for requests they do not wish
616 * to handle. This operation is mandatory.
618 * \param[in,out] policy The policy we're getting resources for.
619 * \param[in,out] nrq The request we are getting resources for.
620 * \param[in] parent The parent resource of the resource being
621 * requested; set to NULL if none.
622 * \param[out] resp The resource is to be returned here; the
623 * fallback policy in an NRS head should
624 * \e always return a non-NULL pointer value.
625 * \param[in] moving_req When set, signifies that this is an attempt
626 * to obtain resources for a request being moved
627 * to the high-priority NRS head by
628 * ldlm_lock_reorder_req().
629 * This implies two things:
630 * 1. We are under obd_export::exp_rpc_lock and
631 * so should not sleep.
632 * 2. We should not perform non-idempotent or can
633 * skip performing idempotent operations that
634 * were carried out when resources were first
635 * taken for the request when it was initialized
636 * in ptlrpc_nrs_req_initialize().
638 * \retval 0, +ve The level of the returned resource in the resource
639 * hierarchy; currently only 0 (for a non-leaf resource)
640 * and 1 (for a leaf resource) are supported by the
644 * \see ptlrpc_nrs_req_initialize()
645 * \see ptlrpc_nrs_hpreq_add_nolock()
646 * \see ptlrpc_nrs_req_hp_move()
648 int (*op_res_get) (struct ptlrpc_nrs_policy *policy,
649 struct ptlrpc_nrs_request *nrq,
650 const struct ptlrpc_nrs_resource *parent,
651 struct ptlrpc_nrs_resource **resp,
654 * Called when releasing references taken for resources in the resource
655 * hierarchy for the request; this operation is optional.
657 * \param[in,out] policy The policy the resource belongs to
658 * \param[in] res The resource to be freed
660 * \see ptlrpc_nrs_req_finalize()
661 * \see ptlrpc_nrs_hpreq_add_nolock()
662 * \see ptlrpc_nrs_req_hp_move()
664 void (*op_res_put) (struct ptlrpc_nrs_policy *policy,
665 const struct ptlrpc_nrs_resource *res);
668 * Obtains a request for handling from the policy, and optionally
669 * removes the request from the policy; this operation is mandatory.
671 * \param[in,out] policy The policy to poll
672 * \param[in] peek When set, signifies that we just want to
673 * examine the request, and not handle it, so the
674 * request is not removed from the policy.
675 * \param[in] force When set, it will force a policy to return a
676 * request if it has one queued.
678 * \retval NULL No request available for handling
679 * \retval valid-pointer The request polled for handling
681 * \see ptlrpc_nrs_req_get_nolock()
683 struct ptlrpc_nrs_request *
684 (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek,
687 * Called when attempting to add a request to a policy for later
688 * handling; this operation is mandatory.
690 * \param[in,out] policy The policy on which to enqueue \a nrq
691 * \param[in,out] nrq The request to enqueue
696 * \see ptlrpc_nrs_req_add_nolock()
698 int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy,
699 struct ptlrpc_nrs_request *nrq);
701 * Removes a request from the policy's set of pending requests. Normally
702 * called after a request has been polled successfully from the policy
703 * for handling; this operation is mandatory.
705 * \param[in,out] policy The policy the request \a nrq belongs to
706 * \param[in,out] nrq The request to dequeue
708 * \see ptlrpc_nrs_req_del_nolock()
710 void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy,
711 struct ptlrpc_nrs_request *nrq);
713 * Called after the request being carried out. Could be used for
714 * job/resource control; this operation is optional.
716 * \param[in,out] policy The policy which is stopping to handle request
718 * \param[in,out] nrq The request
720 * \pre assert_spin_locked(&svcpt->scp_req_lock)
722 * \see ptlrpc_nrs_req_stop_nolock()
724 void (*op_req_stop) (struct ptlrpc_nrs_policy *policy,
725 struct ptlrpc_nrs_request *nrq);
727 * Registers the policy's lprocfs interface with a PTLRPC service.
729 * \param[in] svc The service
734 int (*op_lprocfs_init) (struct ptlrpc_service *svc);
736 * Unegisters the policy's lprocfs interface with a PTLRPC service.
738 * In cases of failed policy registration in
739 * \e ptlrpc_nrs_policy_register(), this function may be called for a
740 * service which has not registered the policy successfully, so
741 * implementations of this method should make sure their operations are
742 * safe in such cases.
744 * \param[in] svc The service
746 void (*op_lprocfs_fini) (struct ptlrpc_service *svc);
752 enum nrs_policy_flags {
754 * Fallback policy, use this flag only on a single supported policy per
755 * service. The flag cannot be used on policies that use
756 * \e PTLRPC_NRS_FL_REG_EXTERN
758 PTLRPC_NRS_FL_FALLBACK = (1 << 0),
760 * Start policy immediately after registering.
762 PTLRPC_NRS_FL_REG_START = (1 << 1),
764 * This is a policy registering from a module different to the one NRS
765 * core ships in (currently ptlrpc).
767 PTLRPC_NRS_FL_REG_EXTERN = (1 << 2),
773 * Denotes whether an NRS instance is for handling normal or high-priority
774 * RPCs, or whether an operation pertains to one or both of the NRS instances
777 enum ptlrpc_nrs_queue_type {
778 PTLRPC_NRS_QUEUE_REG = (1 << 0),
779 PTLRPC_NRS_QUEUE_HP = (1 << 1),
780 PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP)
786 * A PTLRPC service has at least one NRS head instance for handling normal
787 * priority RPCs, and may optionally have a second NRS head instance for
788 * handling high-priority RPCs. Each NRS head maintains a list of available
789 * policies, of which one and only one policy is acting as the fallback policy,
790 * and optionally a different policy may be acting as the primary policy. For
791 * all RPCs handled by this NRS head instance, NRS core will first attempt to
792 * enqueue the RPC using the primary policy (if any). The fallback policy is
793 * used in the following cases:
794 * - when there was no primary policy in the
795 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
797 * - when the primary policy that was at the
798 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
799 * RPC was initialized, denoted it did not wish, or for some other reason was
800 * not able to handle the request, by returning a non-valid NRS resource
802 * - when the primary policy that was at the
803 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
804 * RPC was initialized, fails later during the request enqueueing stage.
806 * \see nrs_resource_get_safe()
807 * \see nrs_request_enqueue()
811 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
813 * List of registered policies
815 struct list_head nrs_policy_list;
817 * List of policies with queued requests. Policies that have any
818 * outstanding requests are queued here, and this list is queried
819 * in a round-robin manner from NRS core when obtaining a request
820 * for handling. This ensures that requests from policies that at some
821 * point transition away from the
822 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
824 struct list_head nrs_policy_queued;
826 * Service partition for this NRS head
828 struct ptlrpc_service_part *nrs_svcpt;
830 * Primary policy, which is the preferred policy for handling RPCs
832 struct ptlrpc_nrs_policy *nrs_policy_primary;
834 * Fallback policy, which is the backup policy for handling RPCs
836 struct ptlrpc_nrs_policy *nrs_policy_fallback;
838 * This NRS head handles either HP or regular requests
840 enum ptlrpc_nrs_queue_type nrs_queue_type;
842 * # queued requests from all policies in this NRS head
844 unsigned long nrs_req_queued;
846 * # scheduled requests from all policies in this NRS head
848 unsigned long nrs_req_started;
850 * # policies on this NRS
852 unsigned nrs_num_pols;
854 * This NRS head is in progress of starting a policy
856 unsigned nrs_policy_starting:1;
858 * In progress of shutting down the whole NRS head; used during
861 unsigned nrs_stopping:1;
864 #define NRS_POL_NAME_MAX 16
866 struct ptlrpc_nrs_pol_desc;
869 * Service compatibility predicate; this determines whether a policy is adequate
870 * for handling RPCs of a particular PTLRPC service.
872 * XXX:This should give the same result during policy registration and
873 * unregistration, and for all partitions of a service; so the result should not
874 * depend on temporal service or other properties, that may influence the
877 typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc,
878 const struct ptlrpc_nrs_pol_desc *desc);
880 struct ptlrpc_nrs_pol_conf {
882 * Human-readable policy name
884 char nc_name[NRS_POL_NAME_MAX];
886 * NRS operations for this policy
888 const struct ptlrpc_nrs_pol_ops *nc_ops;
890 * Service compatibility predicate
892 nrs_pol_desc_compat_t nc_compat;
894 * Set for policies that support a single ptlrpc service, i.e. ones that
895 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
896 * depicts the name of the single service that such policies are
899 const char *nc_compat_svc_name;
901 * Owner module for this policy descriptor; policies registering from a
902 * different module to the one the NRS framework is held within
903 * (currently ptlrpc), should set this field to THIS_MODULE.
905 struct module *nc_owner;
907 * Policy registration flags; a bitmask of \e nrs_policy_flags
913 * NRS policy registering descriptor
915 * Is used to hold a description of a policy that can be passed to NRS core in
916 * order to register the policy with NRS heads in different PTLRPC services.
918 struct ptlrpc_nrs_pol_desc {
920 * Human-readable policy name
922 char pd_name[NRS_POL_NAME_MAX];
924 * Link into nrs_core::nrs_policies
926 struct list_head pd_list;
928 * NRS operations for this policy
930 const struct ptlrpc_nrs_pol_ops *pd_ops;
932 * Service compatibility predicate
934 nrs_pol_desc_compat_t pd_compat;
936 * Set for policies that are compatible with only one PTLRPC service.
938 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
940 const char *pd_compat_svc_name;
942 * Owner module for this policy descriptor.
944 * We need to hold a reference to the module whenever we might make use
945 * of any of the module's contents, i.e.
946 * - If one or more instances of the policy are at a state where they
947 * might be handling a request, i.e.
948 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
949 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
950 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
951 * is taken on the module when
952 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
953 * becomes 0, so that we hold only one reference to the module maximum
956 * We do not need to hold a reference to the module, even though we
957 * might use code and data from the module, in the following cases:
958 * - During external policy registration, because this should happen in
959 * the module's init() function, in which case the module is safe from
960 * removal because a reference is being held on the module by the
961 * kernel, and iirc kmod (and I guess module-init-tools also) will
962 * serialize any racing processes properly anyway.
963 * - During external policy unregistration, because this should happen
964 * in a module's exit() function, and any attempts to start a policy
965 * instance would need to take a reference on the module, and this is
966 * not possible once we have reached the point where the exit()
968 * - During service registration and unregistration, as service setup
969 * and cleanup, and policy registration, unregistration and policy
970 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
971 * as long as users adhere to the convention of registering policies
972 * in init() and unregistering them in module exit() functions, there
973 * should not be a race between these operations.
974 * - During any policy-specific lprocfs operations, because a reference
975 * is held by the kernel on a proc entry that has been entered by a
976 * syscall, so as long as proc entries are removed during unregistration time,
977 * then unregistration and lprocfs operations will be properly
980 struct module *pd_owner;
982 * Bitmask of \e nrs_policy_flags
986 * # of references on this descriptor
994 * Policies transition from one state to the other during their lifetime
996 enum ptlrpc_nrs_pol_state {
998 * Not a valid policy state.
1000 NRS_POL_STATE_INVALID,
1002 * Policies are at this state either at the start of their life, or
1003 * transition here when the user selects a different policy to act
1004 * as the primary one.
1006 NRS_POL_STATE_STOPPED,
1008 * Policy is progress of stopping
1010 NRS_POL_STATE_STOPPING,
1012 * Policy is in progress of starting
1014 NRS_POL_STATE_STARTING,
1016 * A policy is in this state in two cases:
1017 * - it is the fallback policy, which is always in this state.
1018 * - it has been activated by the user; i.e. it is the primary policy,
1020 NRS_POL_STATE_STARTED,
1024 * NRS policy information
1026 * Used for obtaining information for the status of a policy via lprocfs
1028 struct ptlrpc_nrs_pol_info {
1032 char pi_name[NRS_POL_NAME_MAX];
1034 * Current policy state
1036 enum ptlrpc_nrs_pol_state pi_state;
1038 * # RPCs enqueued for later dispatching by the policy
1042 * # RPCs started for dispatch by the policy
1044 long pi_req_started;
1046 * Is this a fallback policy?
1048 unsigned pi_fallback:1;
1054 * There is one instance of this for each policy in each NRS head of each
1055 * PTLRPC service partition.
1057 struct ptlrpc_nrs_policy {
1059 * Linkage into the NRS head's list of policies,
1060 * ptlrpc_nrs:nrs_policy_list
1062 struct list_head pol_list;
1064 * Linkage into the NRS head's list of policies with enqueued
1065 * requests ptlrpc_nrs:nrs_policy_queued
1067 struct list_head pol_list_queued;
1069 * Current state of this policy
1071 enum ptlrpc_nrs_pol_state pol_state;
1073 * Bitmask of nrs_policy_flags
1077 * # RPCs enqueued for later dispatching by the policy
1079 long pol_req_queued;
1081 * # RPCs started for dispatch by the policy
1083 long pol_req_started;
1085 * Usage Reference count taken on the policy instance
1089 * The NRS head this policy has been created at
1091 struct ptlrpc_nrs *pol_nrs;
1093 * Private policy data; varies by policy type
1097 * Policy descriptor for this policy instance.
1099 struct ptlrpc_nrs_pol_desc *pol_desc;
1105 * Resources are embedded into two types of NRS entities:
1106 * - Inside NRS policies, in the policy's private data in
1107 * ptlrpc_nrs_policy::pol_private
1108 * - In objects that act as prime-level scheduling entities in different NRS
1109 * policies; e.g. on a policy that performs round robin or similar order
1110 * scheduling across client NIDs, there would be one NRS resource per unique
1111 * client NID. On a policy which performs round robin scheduling across
1112 * backend filesystem objects, there would be one resource associated with
1113 * each of the backend filesystem objects partaking in the scheduling
1114 * performed by the policy.
1116 * NRS resources share a parent-child relationship, in which resources embedded
1117 * in policy instances are the parent entities, with all scheduling entities
1118 * a policy schedules across being the children, thus forming a simple resource
1119 * hierarchy. This hierarchy may be extended with one or more levels in the
1120 * future if the ability to have more than one primary policy is added.
1122 * Upon request initialization, references to the then active NRS policies are
1123 * taken and used to later handle the dispatching of the request with one of
1126 * \see nrs_resource_get_safe()
1127 * \see ptlrpc_nrs_req_add()
1129 struct ptlrpc_nrs_resource {
1131 * This NRS resource's parent; is NULL for resources embedded in NRS
1132 * policy instances; i.e. those are top-level ones.
1134 struct ptlrpc_nrs_resource *res_parent;
1136 * The policy associated with this resource.
1138 struct ptlrpc_nrs_policy *res_policy;
1151 * This policy is a logical wrapper around previous, non-NRS functionality.
1152 * It dispatches RPCs in the same order as they arrive from the network. This
1153 * policy is currently used as the fallback policy, and the only enabled policy
1154 * on all NRS heads of all PTLRPC service partitions.
1159 * Private data structure for the FIFO policy
1161 struct nrs_fifo_head {
1163 * Resource object for policy instance.
1165 struct ptlrpc_nrs_resource fh_res;
1167 * List of queued requests.
1169 struct list_head fh_list;
1171 * For debugging purposes.
1176 struct nrs_fifo_req {
1177 struct list_head fr_list;
1186 * Instances of this object exist embedded within ptlrpc_request; the main
1187 * purpose of this object is to hold references to the request's resources
1188 * for the lifetime of the request, and to hold properties that policies use
1189 * use for determining the request's scheduling priority.
1191 struct ptlrpc_nrs_request {
1193 * The request's resource hierarchy.
1195 struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX];
1197 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1198 * policy that was used to enqueue the request.
1200 * \see nrs_request_enqueue()
1202 unsigned nr_res_idx;
1203 unsigned nr_initialized:1;
1204 unsigned nr_enqueued:1;
1205 unsigned nr_started:1;
1206 unsigned nr_finalized:1;
1209 * Policy-specific fields, used for determining a request's scheduling
1210 * priority, and other supporting functionality.
1214 * Fields for the FIFO policy
1216 struct nrs_fifo_req fifo;
1219 * Externally-registering policies may want to use this to allocate
1220 * their own request properties.
1228 * Basic request prioritization operations structure.
1229 * The whole idea is centered around locks and RPCs that might affect locks.
1230 * When a lock is contended we try to give priority to RPCs that might lead
1231 * to fastest release of that lock.
1232 * Currently only implemented for OSTs only in a way that makes all
1233 * IO and truncate RPCs that are coming from a locked region where a lock is
1234 * contended a priority over other requests.
1236 struct ptlrpc_hpreq_ops {
1238 * Check if the lock handle of the given lock is the same as
1239 * taken from the request.
1241 int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *);
1243 * Check if the request is a high priority one.
1245 int (*hpreq_check)(struct ptlrpc_request *);
1247 * Called after the request has been handled.
1249 void (*hpreq_fini)(struct ptlrpc_request *);
1253 * Represents remote procedure call.
1255 * This is a staple structure used by everybody wanting to send a request
1258 struct ptlrpc_request {
1259 /* Request type: one of PTL_RPC_MSG_* */
1261 /** Result of request processing */
1264 * Linkage item through which this request is included into
1265 * sending/delayed lists on client and into rqbd list on server
1267 struct list_head rq_list;
1269 * Server side list of incoming unserved requests sorted by arrival
1270 * time. Traversed from time to time to notice about to expire
1271 * requests and sent back "early replies" to clients to let them
1272 * know server is alive and well, just very busy to service their
1275 struct list_head rq_timed_list;
1276 /** server-side history, used for debugging purposes. */
1277 struct list_head rq_history_list;
1278 /** server-side per-export list */
1279 struct list_head rq_exp_list;
1280 /** server-side hp handlers */
1281 struct ptlrpc_hpreq_ops *rq_ops;
1283 /** initial thread servicing this request */
1284 struct ptlrpc_thread *rq_svc_thread;
1286 /** history sequence # */
1287 __u64 rq_history_seq;
1291 /** stub for NRS request */
1292 struct ptlrpc_nrs_request rq_nrq;
1294 /** the index of service's srv_at_array into which request is linked */
1296 /** Lock to protect request flags and some other important bits, like
1300 /** client-side flags are serialized by rq_lock */
1301 unsigned int rq_intr:1, rq_replied:1, rq_err:1,
1302 rq_timedout:1, rq_resend:1, rq_restart:1,
1304 * when ->rq_replay is set, request is kept by the client even
1305 * after server commits corresponding transaction. This is
1306 * used for operations that require sequence of multiple
1307 * requests to be replayed. The only example currently is file
1308 * open/close. When last request in such a sequence is
1309 * committed, ->rq_replay is cleared on all requests in the
1313 rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1,
1314 rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1,
1316 rq_req_unlink:1, rq_reply_unlink:1,
1317 rq_memalloc:1, /* req originated from "kswapd" */
1318 /* server-side flags */
1319 rq_packed_final:1, /* packed final reply */
1320 rq_hp:1, /* high priority RPC */
1321 rq_at_linked:1, /* link into service's srv_at_array */
1322 rq_reply_truncate:1,
1324 /* whether the "rq_set" is a valid one */
1326 rq_generation_set:1,
1327 /* do not resend request on -EINPROGRESS */
1328 rq_no_retry_einprogress:1,
1329 /* allow the req to be sent if the import is in recovery
1333 unsigned int rq_nr_resend;
1335 enum rq_phase rq_phase; /* one of RQ_PHASE_* */
1336 enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */
1337 atomic_t rq_refcount;/* client-side refcount for SENT race,
1338 server-side refcount for multiple replies */
1340 /** Portal to which this request would be sent */
1341 short rq_request_portal; /* XXX FIXME bug 249 */
1342 /** Portal where to wait for reply and where reply would be sent */
1343 short rq_reply_portal; /* XXX FIXME bug 249 */
1347 * !rq_truncate : # reply bytes actually received,
1348 * rq_truncate : required repbuf_len for resend
1350 int rq_nob_received;
1351 /** Request length */
1355 /** Request message - what client sent */
1356 struct lustre_msg *rq_reqmsg;
1357 /** Reply message - server response */
1358 struct lustre_msg *rq_repmsg;
1359 /** Transaction number */
1364 * List item to for replay list. Not yet committed requests get linked
1366 * Also see \a rq_replay comment above.
1368 struct list_head rq_replay_list;
1371 * security and encryption data
1373 struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */
1374 struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */
1375 struct list_head rq_ctx_chain; /**< link to waited ctx */
1377 struct sptlrpc_flavor rq_flvr; /**< for client & server */
1378 enum lustre_sec_part rq_sp_from;
1380 /* client/server security flags */
1382 rq_ctx_init:1, /* context initiation */
1383 rq_ctx_fini:1, /* context destroy */
1384 rq_bulk_read:1, /* request bulk read */
1385 rq_bulk_write:1, /* request bulk write */
1386 /* server authentication flags */
1387 rq_auth_gss:1, /* authenticated by gss */
1388 rq_auth_remote:1, /* authed as remote user */
1389 rq_auth_usr_root:1, /* authed as root */
1390 rq_auth_usr_mdt:1, /* authed as mdt */
1391 rq_auth_usr_ost:1, /* authed as ost */
1392 /* security tfm flags */
1395 /* doesn't expect reply FIXME */
1397 rq_pill_init:1; /* pill initialized */
1399 uid_t rq_auth_uid; /* authed uid */
1400 uid_t rq_auth_mapped_uid; /* authed uid mapped to */
1402 /* (server side), pointed directly into req buffer */
1403 struct ptlrpc_user_desc *rq_user_desc;
1405 /* various buffer pointers */
1406 struct lustre_msg *rq_reqbuf; /* req wrapper */
1407 char *rq_repbuf; /* rep buffer */
1408 struct lustre_msg *rq_repdata; /* rep wrapper msg */
1409 struct lustre_msg *rq_clrbuf; /* only in priv mode */
1410 int rq_reqbuf_len; /* req wrapper buf len */
1411 int rq_reqdata_len; /* req wrapper msg len */
1412 int rq_repbuf_len; /* rep buffer len */
1413 int rq_repdata_len; /* rep wrapper msg len */
1414 int rq_clrbuf_len; /* only in priv mode */
1415 int rq_clrdata_len; /* only in priv mode */
1417 /** early replies go to offset 0, regular replies go after that */
1418 unsigned int rq_reply_off;
1422 /** Fields that help to see if request and reply were swabbed or not */
1423 __u32 rq_req_swab_mask;
1424 __u32 rq_rep_swab_mask;
1426 /** What was import generation when this request was sent */
1427 int rq_import_generation;
1428 enum lustre_imp_state rq_send_state;
1430 /** how many early replies (for stats) */
1433 /** client+server request */
1434 lnet_handle_md_t rq_req_md_h;
1435 struct ptlrpc_cb_id rq_req_cbid;
1436 /** optional time limit for send attempts */
1437 long rq_delay_limit;
1438 /** time request was first queued */
1439 unsigned long rq_queued_time;
1441 /* server-side... */
1442 /** request arrival time */
1443 struct timeval rq_arrival_time;
1444 /** separated reply state */
1445 struct ptlrpc_reply_state *rq_reply_state;
1446 /** incoming request buffer */
1447 struct ptlrpc_request_buffer_desc *rq_rqbd;
1449 /** client-only incoming reply */
1450 lnet_handle_md_t rq_reply_md_h;
1451 wait_queue_head_t rq_reply_waitq;
1452 struct ptlrpc_cb_id rq_reply_cbid;
1456 /** Peer description (the other side) */
1457 lnet_process_id_t rq_peer;
1458 /** Server-side, export on which request was received */
1459 struct obd_export *rq_export;
1460 /** Client side, import where request is being sent */
1461 struct obd_import *rq_import;
1463 /** Replay callback, called after request is replayed at recovery */
1464 void (*rq_replay_cb)(struct ptlrpc_request *);
1466 * Commit callback, called when request is committed and about to be
1469 void (*rq_commit_cb)(struct ptlrpc_request *);
1470 /** Opaq data for replay and commit callbacks. */
1473 /** For bulk requests on client only: bulk descriptor */
1474 struct ptlrpc_bulk_desc *rq_bulk;
1476 /** client outgoing req */
1478 * when request/reply sent (secs), or time when request should be sent
1481 /** time for request really sent out */
1482 time_t rq_real_sent;
1484 /** when request must finish. volatile
1485 * so that servers' early reply updates to the deadline aren't
1486 * kept in per-cpu cache */
1487 volatile time_t rq_deadline;
1488 /** when req reply unlink must finish. */
1489 time_t rq_reply_deadline;
1490 /** when req bulk unlink must finish. */
1491 time_t rq_bulk_deadline;
1493 * service time estimate (secs)
1494 * If the requestsis not served by this time, it is marked as timed out.
1498 /** Multi-rpc bits */
1499 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1500 wait_queue_head_t rq_set_waitq;
1501 /** Link item for request set lists */
1502 struct list_head rq_set_chain;
1503 /** Link back to the request set */
1504 struct ptlrpc_request_set *rq_set;
1505 /** Async completion handler, called when reply is received */
1506 ptlrpc_interpterer_t rq_interpret_reply;
1507 /** Async completion context */
1508 union ptlrpc_async_args rq_async_args;
1510 /** Pool if request is from preallocated list */
1511 struct ptlrpc_request_pool *rq_pool;
1513 struct lu_context rq_session;
1514 struct lu_context rq_recov_session;
1516 /** request format description */
1517 struct req_capsule rq_pill;
1521 * Call completion handler for rpc if any, return it's status or original
1522 * rc if there was no handler defined for this request.
1524 static inline int ptlrpc_req_interpret(const struct lu_env *env,
1525 struct ptlrpc_request *req, int rc)
1527 if (req->rq_interpret_reply != NULL) {
1528 req->rq_status = req->rq_interpret_reply(env, req,
1529 &req->rq_async_args,
1531 return req->rq_status;
1539 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf);
1540 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf);
1541 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req);
1542 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy,
1543 struct ptlrpc_nrs_pol_info *info);
1546 * Can the request be moved from the regular NRS head to the high-priority NRS
1547 * head (of the same PTLRPC service partition), if any?
1549 * For a reliable result, this should be checked under svcpt->scp_req lock.
1551 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req)
1553 struct ptlrpc_nrs_request *nrq = &req->rq_nrq;
1556 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1557 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1558 * to make sure it has not been scheduled yet (analogous to previous
1559 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1561 return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp;
1566 * Returns 1 if request buffer at offset \a index was already swabbed
1568 static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index)
1570 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1571 return req->rq_req_swab_mask & (1 << index);
1575 * Returns 1 if request reply buffer at offset \a index was already swabbed
1577 static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index)
1579 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1580 return req->rq_rep_swab_mask & (1 << index);
1584 * Returns 1 if request needs to be swabbed into local cpu byteorder
1586 static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req)
1588 return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1592 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1594 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req)
1596 return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1600 * Mark request buffer at offset \a index that it was already swabbed
1602 static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index)
1604 LASSERT(index < sizeof(req->rq_req_swab_mask) * 8);
1605 LASSERT((req->rq_req_swab_mask & (1 << index)) == 0);
1606 req->rq_req_swab_mask |= 1 << index;
1610 * Mark request reply buffer at offset \a index that it was already swabbed
1612 static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index)
1614 LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8);
1615 LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0);
1616 req->rq_rep_swab_mask |= 1 << index;
1620 * Convert numerical request phase value \a phase into text string description
1622 static inline const char *
1623 ptlrpc_phase2str(enum rq_phase phase)
1632 case RQ_PHASE_INTERPRET:
1634 case RQ_PHASE_COMPLETE:
1636 case RQ_PHASE_UNREGISTERING:
1637 return "Unregistering";
1644 * Convert numerical request phase of the request \a req into text stringi
1647 static inline const char *
1648 ptlrpc_rqphase2str(struct ptlrpc_request *req)
1650 return ptlrpc_phase2str(req->rq_phase);
1654 * Debugging functions and helpers to print request structure into debug log
1657 /* Spare the preprocessor, spoil the bugs. */
1658 #define FLAG(field, str) (field ? str : "")
1660 /** Convert bit flags into a string */
1661 #define DEBUG_REQ_FLAGS(req) \
1662 ptlrpc_rqphase2str(req), \
1663 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1664 FLAG(req->rq_err, "E"), \
1665 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1666 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1667 FLAG(req->rq_no_resend, "N"), \
1668 FLAG(req->rq_waiting, "W"), \
1669 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1670 FLAG(req->rq_committed, "M")
1672 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1674 void _debug_req(struct ptlrpc_request *req,
1675 struct libcfs_debug_msg_data *data, const char *fmt, ...)
1679 * Helper that decides if we need to print request according to current debug
1682 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1684 CFS_CHECK_STACK(msgdata, mask, cdls); \
1686 if (((mask) & D_CANTMASK) != 0 || \
1687 ((libcfs_debug & (mask)) != 0 && \
1688 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1689 _debug_req((req), msgdata, fmt, ##a); \
1693 * This is the debug print function you need to use to print request structure
1694 * content into lustre debug log.
1695 * for most callers (level is a constant) this is resolved at compile time */
1696 #define DEBUG_REQ(level, req, fmt, args...) \
1698 if ((level) & (D_ERROR | D_WARNING)) { \
1699 static struct cfs_debug_limit_state cdls; \
1700 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1701 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1703 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1704 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1710 * Structure that defines a single page of a bulk transfer
1712 struct ptlrpc_bulk_page {
1713 /** Linkage to list of pages in a bulk */
1714 struct list_head bp_link;
1716 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1719 /** offset within a page */
1721 /** The page itself */
1722 struct page *bp_page;
1725 #define BULK_GET_SOURCE 0
1726 #define BULK_PUT_SINK 1
1727 #define BULK_GET_SINK 2
1728 #define BULK_PUT_SOURCE 3
1731 * Definition of bulk descriptor.
1732 * Bulks are special "Two phase" RPCs where initial request message
1733 * is sent first and it is followed bt a transfer (o receiving) of a large
1734 * amount of data to be settled into pages referenced from the bulk descriptors.
1735 * Bulks transfers (the actual data following the small requests) are done
1736 * on separate LNet portals.
1737 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1738 * Another user is readpage for MDT.
1740 struct ptlrpc_bulk_desc {
1741 /** completed with failure */
1742 unsigned long bd_failure:1;
1743 /** {put,get}{source,sink} */
1744 unsigned long bd_type:2;
1746 unsigned long bd_registered:1;
1747 /** For serialization with callback */
1749 /** Import generation when request for this bulk was sent */
1750 int bd_import_generation;
1751 /** LNet portal for this bulk */
1753 /** Server side - export this bulk created for */
1754 struct obd_export *bd_export;
1755 /** Client side - import this bulk was sent on */
1756 struct obd_import *bd_import;
1757 /** Back pointer to the request */
1758 struct ptlrpc_request *bd_req;
1759 wait_queue_head_t bd_waitq; /* server side only WQ */
1760 int bd_iov_count; /* # entries in bd_iov */
1761 int bd_max_iov; /* allocated size of bd_iov */
1762 int bd_nob; /* # bytes covered */
1763 int bd_nob_transferred; /* # bytes GOT/PUT */
1767 struct ptlrpc_cb_id bd_cbid; /* network callback info */
1768 lnet_nid_t bd_sender; /* stash event::sender */
1769 int bd_md_count; /* # valid entries in bd_mds */
1770 int bd_md_max_brw; /* max entries in bd_mds */
1771 /** array of associated MDs */
1772 lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT];
1775 * encrypt iov, size is either 0 or bd_iov_count.
1777 lnet_kiov_t *bd_enc_iov;
1779 lnet_kiov_t bd_iov[0];
1783 SVC_STOPPED = 1 << 0,
1784 SVC_STOPPING = 1 << 1,
1785 SVC_STARTING = 1 << 2,
1786 SVC_RUNNING = 1 << 3,
1788 SVC_SIGNAL = 1 << 5,
1791 #define PTLRPC_THR_NAME_LEN 32
1793 * Definition of server service thread structure
1795 struct ptlrpc_thread {
1797 * List of active threads in svc->srv_threads
1799 struct list_head t_link;
1801 * thread-private data (preallocated memory)
1806 * service thread index, from ptlrpc_start_threads
1810 * service thread pid
1814 * put watchdog in the structure per thread b=14840
1816 * Lustre watchdog is removed for client in the hope
1817 * of a generic watchdog can be merged in kernel.
1818 * When that happens, we should add below back.
1820 * struct lc_watchdog *t_watchdog;
1823 * the svc this thread belonged to b=18582
1825 struct ptlrpc_service_part *t_svcpt;
1826 wait_queue_head_t t_ctl_waitq;
1827 struct lu_env *t_env;
1828 char t_name[PTLRPC_THR_NAME_LEN];
1831 static inline int thread_is_init(struct ptlrpc_thread *thread)
1833 return thread->t_flags == 0;
1836 static inline int thread_is_stopped(struct ptlrpc_thread *thread)
1838 return !!(thread->t_flags & SVC_STOPPED);
1841 static inline int thread_is_stopping(struct ptlrpc_thread *thread)
1843 return !!(thread->t_flags & SVC_STOPPING);
1846 static inline int thread_is_starting(struct ptlrpc_thread *thread)
1848 return !!(thread->t_flags & SVC_STARTING);
1851 static inline int thread_is_running(struct ptlrpc_thread *thread)
1853 return !!(thread->t_flags & SVC_RUNNING);
1856 static inline int thread_is_event(struct ptlrpc_thread *thread)
1858 return !!(thread->t_flags & SVC_EVENT);
1861 static inline int thread_is_signal(struct ptlrpc_thread *thread)
1863 return !!(thread->t_flags & SVC_SIGNAL);
1866 static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags)
1868 thread->t_flags &= ~flags;
1871 static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags)
1873 thread->t_flags = flags;
1876 static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags)
1878 thread->t_flags |= flags;
1881 static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread,
1884 if (thread->t_flags & flags) {
1885 thread->t_flags &= ~flags;
1892 * Request buffer descriptor structure.
1893 * This is a structure that contains one posted request buffer for service.
1894 * Once data land into a buffer, event callback creates actual request and
1895 * notifies wakes one of the service threads to process new incoming request.
1896 * More than one request can fit into the buffer.
1898 struct ptlrpc_request_buffer_desc {
1899 /** Link item for rqbds on a service */
1900 struct list_head rqbd_list;
1901 /** History of requests for this buffer */
1902 struct list_head rqbd_reqs;
1903 /** Back pointer to service for which this buffer is registered */
1904 struct ptlrpc_service_part *rqbd_svcpt;
1905 /** LNet descriptor */
1906 lnet_handle_md_t rqbd_md_h;
1908 /** The buffer itself */
1910 struct ptlrpc_cb_id rqbd_cbid;
1912 * This "embedded" request structure is only used for the
1913 * last request to fit into the buffer
1915 struct ptlrpc_request rqbd_req;
1918 typedef int (*svc_handler_t)(struct ptlrpc_request *req);
1920 struct ptlrpc_service_ops {
1922 * if non-NULL called during thread creation (ptlrpc_start_thread())
1923 * to initialize service specific per-thread state.
1925 int (*so_thr_init)(struct ptlrpc_thread *thr);
1927 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1928 * destruct state created by ->srv_init().
1930 void (*so_thr_done)(struct ptlrpc_thread *thr);
1932 * Handler function for incoming requests for this service
1934 int (*so_req_handler)(struct ptlrpc_request *req);
1936 * function to determine priority of the request, it's called
1937 * on every new request
1939 int (*so_hpreq_handler)(struct ptlrpc_request *);
1941 * service-specific print fn
1943 void (*so_req_printer)(void *, struct ptlrpc_request *);
1946 #ifndef __cfs_cacheline_aligned
1947 /* NB: put it here for reducing patche dependence */
1948 # define __cfs_cacheline_aligned
1952 * How many high priority requests to serve before serving one normal
1955 #define PTLRPC_SVC_HP_RATIO 10
1958 * Definition of PortalRPC service.
1959 * The service is listening on a particular portal (like tcp port)
1960 * and perform actions for a specific server like IO service for OST
1961 * or general metadata service for MDS.
1963 struct ptlrpc_service {
1964 /** serialize /proc operations */
1965 spinlock_t srv_lock;
1966 /** most often accessed fields */
1967 /** chain thru all services */
1968 struct list_head srv_list;
1969 /** service operations table */
1970 struct ptlrpc_service_ops srv_ops;
1971 /** only statically allocated strings here; we don't clean them */
1973 /** only statically allocated strings here; we don't clean them */
1974 char *srv_thread_name;
1975 /** service thread list */
1976 struct list_head srv_threads;
1977 /** threads # should be created for each partition on initializing */
1978 int srv_nthrs_cpt_init;
1979 /** limit of threads number for each partition */
1980 int srv_nthrs_cpt_limit;
1981 /** Root of /proc dir tree for this service */
1982 struct proc_dir_entry *srv_procroot;
1983 /** Pointer to statistic data for this service */
1984 struct lprocfs_stats *srv_stats;
1985 /** # hp per lp reqs to handle */
1986 int srv_hpreq_ratio;
1987 /** biggest request to receive */
1988 int srv_max_req_size;
1989 /** biggest reply to send */
1990 int srv_max_reply_size;
1991 /** size of individual buffers */
1993 /** # buffers to allocate in 1 group */
1994 int srv_nbuf_per_group;
1995 /** Local portal on which to receive requests */
1996 __u32 srv_req_portal;
1997 /** Portal on the client to send replies to */
1998 __u32 srv_rep_portal;
2000 * Tags for lu_context associated with this thread, see struct
2004 /** soft watchdog timeout multiplier */
2005 int srv_watchdog_factor;
2006 /** under unregister_service */
2007 unsigned srv_is_stopping:1;
2009 /** max # request buffers in history per partition */
2010 int srv_hist_nrqbds_cpt_max;
2011 /** number of CPTs this service bound on */
2013 /** CPTs array this service bound on */
2015 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2017 /** CPT table this service is running over */
2018 struct cfs_cpt_table *srv_cptable;
2020 * partition data for ptlrpc service
2022 struct ptlrpc_service_part *srv_parts[0];
2026 * Definition of PortalRPC service partition data.
2027 * Although a service only has one instance of it right now, but we
2028 * will have multiple instances very soon (instance per CPT).
2030 * it has four locks:
2032 * serialize operations on rqbd and requests waiting for preprocess
2034 * serialize operations active requests sent to this portal
2036 * serialize adaptive timeout stuff
2038 * serialize operations on RS list (reply states)
2040 * We don't have any use-case to take two or more locks at the same time
2041 * for now, so there is no lock order issue.
2043 struct ptlrpc_service_part {
2044 /** back reference to owner */
2045 struct ptlrpc_service *scp_service __cfs_cacheline_aligned;
2046 /* CPT id, reserved */
2048 /** always increasing number */
2050 /** # of starting threads */
2051 int scp_nthrs_starting;
2052 /** # of stopping threads, reserved for shrinking threads */
2053 int scp_nthrs_stopping;
2054 /** # running threads */
2055 int scp_nthrs_running;
2056 /** service threads list */
2057 struct list_head scp_threads;
2060 * serialize the following fields, used for protecting
2061 * rqbd list and incoming requests waiting for preprocess,
2062 * threads starting & stopping are also protected by this lock.
2064 spinlock_t scp_lock __cfs_cacheline_aligned;
2065 /** total # req buffer descs allocated */
2066 int scp_nrqbds_total;
2067 /** # posted request buffers for receiving */
2068 int scp_nrqbds_posted;
2069 /** in progress of allocating rqbd */
2070 int scp_rqbd_allocating;
2071 /** # incoming reqs */
2072 int scp_nreqs_incoming;
2073 /** request buffers to be reposted */
2074 struct list_head scp_rqbd_idle;
2075 /** req buffers receiving */
2076 struct list_head scp_rqbd_posted;
2077 /** incoming reqs */
2078 struct list_head scp_req_incoming;
2079 /** timeout before re-posting reqs, in tick */
2080 long scp_rqbd_timeout;
2082 * all threads sleep on this. This wait-queue is signalled when new
2083 * incoming request arrives and when difficult reply has to be handled.
2085 wait_queue_head_t scp_waitq;
2087 /** request history */
2088 struct list_head scp_hist_reqs;
2089 /** request buffer history */
2090 struct list_head scp_hist_rqbds;
2091 /** # request buffers in history */
2092 int scp_hist_nrqbds;
2093 /** sequence number for request */
2095 /** highest seq culled from history */
2096 __u64 scp_hist_seq_culled;
2099 * serialize the following fields, used for processing requests
2100 * sent to this portal
2102 spinlock_t scp_req_lock __cfs_cacheline_aligned;
2103 /** # reqs in either of the NRS heads below */
2104 /** # reqs being served */
2105 int scp_nreqs_active;
2106 /** # HPreqs being served */
2107 int scp_nhreqs_active;
2108 /** # hp requests handled */
2111 /** NRS head for regular requests */
2112 struct ptlrpc_nrs scp_nrs_reg;
2113 /** NRS head for HP requests; this is only valid for services that can
2114 * handle HP requests */
2115 struct ptlrpc_nrs *scp_nrs_hp;
2120 * serialize the following fields, used for changes on
2123 spinlock_t scp_at_lock __cfs_cacheline_aligned;
2124 /** estimated rpc service time */
2125 struct adaptive_timeout scp_at_estimate;
2126 /** reqs waiting for replies */
2127 struct ptlrpc_at_array scp_at_array;
2128 /** early reply timer */
2129 struct timer_list scp_at_timer;
2131 unsigned long scp_at_checktime;
2132 /** check early replies */
2133 unsigned scp_at_check;
2137 * serialize the following fields, used for processing
2138 * replies for this portal
2140 spinlock_t scp_rep_lock __cfs_cacheline_aligned;
2141 /** all the active replies */
2142 struct list_head scp_rep_active;
2143 /** List of free reply_states */
2144 struct list_head scp_rep_idle;
2145 /** waitq to run, when adding stuff to srv_free_rs_list */
2146 wait_queue_head_t scp_rep_waitq;
2147 /** # 'difficult' replies */
2148 atomic_t scp_nreps_difficult;
2151 #define ptlrpc_service_for_each_part(part, i, svc) \
2153 i < (svc)->srv_ncpts && \
2154 (svc)->srv_parts != NULL && \
2155 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2158 * Declaration of ptlrpcd control structure
2160 struct ptlrpcd_ctl {
2162 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2164 unsigned long pc_flags;
2166 * Thread lock protecting structure fields.
2172 struct completion pc_starting;
2176 struct completion pc_finishing;
2178 * Thread requests set.
2180 struct ptlrpc_request_set *pc_set;
2182 * Thread name used in cfs_daemonize()
2186 * Environment for request interpreters to run in.
2188 struct lu_env pc_env;
2190 * Index of ptlrpcd thread in the array.
2194 * Number of the ptlrpcd's partners.
2198 * Pointer to the array of partners' ptlrpcd_ctl structure.
2200 struct ptlrpcd_ctl **pc_partners;
2202 * Record the partner index to be processed next.
2207 /* Bits for pc_flags */
2208 enum ptlrpcd_ctl_flags {
2210 * Ptlrpc thread start flag.
2212 LIOD_START = 1 << 0,
2214 * Ptlrpc thread stop flag.
2218 * Ptlrpc thread force flag (only stop force so far).
2219 * This will cause aborting any inflight rpcs handled
2220 * by thread if LIOD_STOP is specified.
2222 LIOD_FORCE = 1 << 2,
2224 * This is a recovery ptlrpc thread.
2226 LIOD_RECOVERY = 1 << 3,
2228 * The ptlrpcd is bound to some CPU core.
2237 * Service compatibility function; the policy is compatible with all services.
2239 * \param[in] svc The service the policy is attempting to register with.
2240 * \param[in] desc The policy descriptor
2242 * \retval true The policy is compatible with the service
2244 * \see ptlrpc_nrs_pol_desc::pd_compat()
2246 static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc,
2247 const struct ptlrpc_nrs_pol_desc *desc)
2253 * Service compatibility function; the policy is compatible with only a specific
2254 * service which is identified by its human-readable name at
2255 * ptlrpc_service::srv_name.
2257 * \param[in] svc The service the policy is attempting to register with.
2258 * \param[in] desc The policy descriptor
2260 * \retval false The policy is not compatible with the service
2261 * \retval true The policy is compatible with the service
2263 * \see ptlrpc_nrs_pol_desc::pd_compat()
2265 static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc,
2266 const struct ptlrpc_nrs_pol_desc *desc)
2268 LASSERT(desc->pd_compat_svc_name != NULL);
2269 return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0;
2274 /* ptlrpc/events.c */
2275 extern lnet_handle_eq_t ptlrpc_eq_h;
2276 extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
2277 lnet_process_id_t *peer, lnet_nid_t *self);
2279 * These callbacks are invoked by LNet when something happened to
2283 extern void request_out_callback(lnet_event_t *ev);
2284 extern void reply_in_callback(lnet_event_t *ev);
2285 extern void client_bulk_callback(lnet_event_t *ev);
2286 extern void request_in_callback(lnet_event_t *ev);
2287 extern void reply_out_callback(lnet_event_t *ev);
2290 /* ptlrpc/connection.c */
2291 struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer,
2293 struct obd_uuid *uuid);
2294 int ptlrpc_connection_put(struct ptlrpc_connection *c);
2295 struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *);
2296 int ptlrpc_connection_init(void);
2297 void ptlrpc_connection_fini(void);
2298 extern lnet_pid_t ptl_get_pid(void);
2300 /* ptlrpc/niobuf.c */
2302 * Actual interfacing with LNet to put/get/register/unregister stuff
2306 int ptlrpc_register_bulk(struct ptlrpc_request *req);
2307 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async);
2309 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req)
2311 struct ptlrpc_bulk_desc *desc;
2314 LASSERT(req != NULL);
2315 desc = req->rq_bulk;
2317 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
2318 req->rq_bulk_deadline > get_seconds())
2324 spin_lock(&desc->bd_lock);
2325 rc = desc->bd_md_count;
2326 spin_unlock(&desc->bd_lock);
2330 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2331 #define PTLRPC_REPLY_EARLY 0x02
2332 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags);
2333 int ptlrpc_reply(struct ptlrpc_request *req);
2334 int ptlrpc_send_error(struct ptlrpc_request *req, int difficult);
2335 int ptlrpc_error(struct ptlrpc_request *req);
2336 void ptlrpc_resend_req(struct ptlrpc_request *request);
2337 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req);
2338 int ptl_send_rpc(struct ptlrpc_request *request, int noreply);
2339 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd);
2342 /* ptlrpc/client.c */
2344 * Client-side portals API. Everything to send requests, receive replies,
2345 * request queues, request management, etc.
2348 void ptlrpc_request_committed(struct ptlrpc_request *req, int force);
2350 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
2351 struct ptlrpc_client *);
2352 void ptlrpc_cleanup_client(struct obd_import *imp);
2353 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid);
2355 int ptlrpc_queue_wait(struct ptlrpc_request *req);
2356 int ptlrpc_replay_req(struct ptlrpc_request *req);
2357 int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async);
2358 void ptlrpc_restart_req(struct ptlrpc_request *req);
2359 void ptlrpc_abort_inflight(struct obd_import *imp);
2360 void ptlrpc_cleanup_imp(struct obd_import *imp);
2361 void ptlrpc_abort_set(struct ptlrpc_request_set *set);
2363 struct ptlrpc_request_set *ptlrpc_prep_set(void);
2364 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
2366 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
2367 set_interpreter_func fn, void *data);
2368 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *);
2369 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set);
2370 int ptlrpc_set_wait(struct ptlrpc_request_set *);
2371 int ptlrpc_expired_set(void *data);
2372 void ptlrpc_interrupted_set(void *data);
2373 void ptlrpc_mark_interrupted(struct ptlrpc_request *req);
2374 void ptlrpc_set_destroy(struct ptlrpc_request_set *);
2375 void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *);
2376 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
2377 struct ptlrpc_request *req);
2379 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool);
2380 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq);
2382 struct ptlrpc_request_pool *
2383 ptlrpc_init_rq_pool(int, int,
2384 void (*populate_pool)(struct ptlrpc_request_pool *, int));
2386 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req);
2387 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
2388 const struct req_format *format);
2389 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
2390 struct ptlrpc_request_pool *,
2391 const struct req_format *format);
2392 void ptlrpc_request_free(struct ptlrpc_request *request);
2393 int ptlrpc_request_pack(struct ptlrpc_request *request,
2394 __u32 version, int opcode);
2395 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
2396 const struct req_format *format,
2397 __u32 version, int opcode);
2398 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
2399 __u32 version, int opcode, char **bufs,
2400 struct ptlrpc_cli_ctx *ctx);
2401 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
2402 int opcode, int count, __u32 *lengths,
2404 struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp,
2405 __u32 version, int opcode,
2406 int count, __u32 *lengths, char **bufs,
2407 struct ptlrpc_request_pool *pool);
2408 void ptlrpc_req_finished(struct ptlrpc_request *request);
2409 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request);
2410 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req);
2411 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
2412 unsigned npages, unsigned max_brw,
2413 unsigned type, unsigned portal);
2414 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin);
2415 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk)
2417 __ptlrpc_free_bulk(bulk, 1);
2419 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk)
2421 __ptlrpc_free_bulk(bulk, 0);
2423 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
2424 struct page *page, int pageoffset, int len, int);
2425 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
2426 struct page *page, int pageoffset,
2429 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
2432 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
2433 struct page *page, int pageoffset,
2436 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
2439 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2440 struct obd_import *imp);
2441 __u64 ptlrpc_next_xid(void);
2442 __u64 ptlrpc_sample_next_xid(void);
2443 __u64 ptlrpc_req_xid(struct ptlrpc_request *request);
2445 /* Set of routines to run a function in ptlrpcd context */
2446 void *ptlrpcd_alloc_work(struct obd_import *imp,
2447 int (*cb)(const struct lu_env *, void *), void *data);
2448 void ptlrpcd_destroy_work(void *handler);
2449 int ptlrpcd_queue_work(void *handler);
2452 struct ptlrpc_service_buf_conf {
2453 /* nbufs is buffers # to allocate when growing the pool */
2454 unsigned int bc_nbufs;
2455 /* buffer size to post */
2456 unsigned int bc_buf_size;
2457 /* portal to listed for requests on */
2458 unsigned int bc_req_portal;
2459 /* portal of where to send replies to */
2460 unsigned int bc_rep_portal;
2461 /* maximum request size to be accepted for this service */
2462 unsigned int bc_req_max_size;
2463 /* maximum reply size this service can ever send */
2464 unsigned int bc_rep_max_size;
2467 struct ptlrpc_service_thr_conf {
2468 /* threadname should be 8 characters or less - 6 will be added on */
2470 /* threads increasing factor for each CPU */
2471 unsigned int tc_thr_factor;
2472 /* service threads # to start on each partition while initializing */
2473 unsigned int tc_nthrs_init;
2475 * low water of threads # upper-limit on each partition while running,
2476 * service availability may be impacted if threads number is lower
2477 * than this value. It can be ZERO if the service doesn't require
2478 * CPU affinity or there is only one partition.
2480 unsigned int tc_nthrs_base;
2481 /* "soft" limit for total threads number */
2482 unsigned int tc_nthrs_max;
2483 /* user specified threads number, it will be validated due to
2484 * other members of this structure. */
2485 unsigned int tc_nthrs_user;
2486 /* set NUMA node affinity for service threads */
2487 unsigned int tc_cpu_affinity;
2488 /* Tags for lu_context associated with service thread */
2492 struct ptlrpc_service_cpt_conf {
2493 struct cfs_cpt_table *cc_cptable;
2494 /* string pattern to describe CPTs for a service */
2498 struct ptlrpc_service_conf {
2501 /* soft watchdog timeout multiplifier to print stuck service traces */
2502 unsigned int psc_watchdog_factor;
2503 /* buffer information */
2504 struct ptlrpc_service_buf_conf psc_buf;
2505 /* thread information */
2506 struct ptlrpc_service_thr_conf psc_thr;
2507 /* CPU partition information */
2508 struct ptlrpc_service_cpt_conf psc_cpt;
2509 /* function table */
2510 struct ptlrpc_service_ops psc_ops;
2513 /* ptlrpc/service.c */
2515 * Server-side services API. Register/unregister service, request state
2516 * management, service thread management
2520 void ptlrpc_save_lock(struct ptlrpc_request *req,
2521 struct lustre_handle *lock, int mode, int no_ack);
2522 void ptlrpc_commit_replies(struct obd_export *exp);
2523 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs);
2524 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs);
2525 int ptlrpc_hpreq_handler(struct ptlrpc_request *req);
2526 struct ptlrpc_service *ptlrpc_register_service(
2527 struct ptlrpc_service_conf *conf,
2528 struct proc_dir_entry *proc_entry);
2529 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc);
2531 int ptlrpc_start_threads(struct ptlrpc_service *svc);
2532 int ptlrpc_unregister_service(struct ptlrpc_service *service);
2533 int liblustre_check_services(void *arg);
2534 void ptlrpc_daemonize(char *name);
2535 int ptlrpc_service_health_check(struct ptlrpc_service *);
2536 void ptlrpc_server_drop_request(struct ptlrpc_request *req);
2537 void ptlrpc_request_change_export(struct ptlrpc_request *req,
2538 struct obd_export *export);
2540 int ptlrpc_hr_init(void);
2541 void ptlrpc_hr_fini(void);
2545 /* ptlrpc/import.c */
2550 int ptlrpc_connect_import(struct obd_import *imp);
2551 int ptlrpc_init_import(struct obd_import *imp);
2552 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose);
2553 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
2554 void deuuidify(char *uuid, const char *prefix, char **uuid_start,
2557 /* ptlrpc/pack_generic.c */
2558 int ptlrpc_reconnect_import(struct obd_import *imp);
2562 * ptlrpc msg buffer and swab interface
2566 int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout,
2568 void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout,
2570 int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len);
2571 int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len);
2573 int lustre_msg_check_version(struct lustre_msg *msg, __u32 version);
2574 void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens,
2576 int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count,
2577 __u32 *lens, char **bufs);
2578 int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens,
2580 int lustre_pack_reply_v2(struct ptlrpc_request *req, int count,
2581 __u32 *lens, char **bufs, int flags);
2582 #define LPRFL_EARLY_REPLY 1
2583 int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens,
2584 char **bufs, int flags);
2585 int lustre_shrink_msg(struct lustre_msg *msg, int segment,
2586 unsigned int newlen, int move_data);
2587 void lustre_free_reply_state(struct ptlrpc_reply_state *rs);
2588 int __lustre_unpack_msg(struct lustre_msg *m, int len);
2589 int lustre_msg_hdr_size(__u32 magic, int count);
2590 int lustre_msg_size(__u32 magic, int count, __u32 *lengths);
2591 int lustre_msg_size_v2(int count, __u32 *lengths);
2592 int lustre_packed_msg_size(struct lustre_msg *msg);
2593 int lustre_msg_early_size(void);
2594 void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size);
2595 void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen);
2596 int lustre_msg_buflen(struct lustre_msg *m, int n);
2597 void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len);
2598 int lustre_msg_bufcount(struct lustre_msg *m);
2599 char *lustre_msg_string(struct lustre_msg *m, int n, int max_len);
2600 __u32 lustre_msghdr_get_flags(struct lustre_msg *msg);
2601 void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags);
2602 __u32 lustre_msg_get_flags(struct lustre_msg *msg);
2603 void lustre_msg_add_flags(struct lustre_msg *msg, int flags);
2604 void lustre_msg_set_flags(struct lustre_msg *msg, int flags);
2605 void lustre_msg_clear_flags(struct lustre_msg *msg, int flags);
2606 __u32 lustre_msg_get_op_flags(struct lustre_msg *msg);
2607 void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags);
2608 void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags);
2609 struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg);
2610 __u32 lustre_msg_get_type(struct lustre_msg *msg);
2611 __u32 lustre_msg_get_version(struct lustre_msg *msg);
2612 void lustre_msg_add_version(struct lustre_msg *msg, int version);
2613 __u32 lustre_msg_get_opc(struct lustre_msg *msg);
2614 __u64 lustre_msg_get_last_xid(struct lustre_msg *msg);
2615 __u64 lustre_msg_get_last_committed(struct lustre_msg *msg);
2616 __u64 *lustre_msg_get_versions(struct lustre_msg *msg);
2617 __u64 lustre_msg_get_transno(struct lustre_msg *msg);
2618 __u64 lustre_msg_get_slv(struct lustre_msg *msg);
2619 __u32 lustre_msg_get_limit(struct lustre_msg *msg);
2620 void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv);
2621 void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit);
2622 int lustre_msg_get_status(struct lustre_msg *msg);
2623 __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg);
2624 int lustre_msg_is_v1(struct lustre_msg *msg);
2625 __u32 lustre_msg_get_magic(struct lustre_msg *msg);
2626 __u32 lustre_msg_get_timeout(struct lustre_msg *msg);
2627 __u32 lustre_msg_get_service_time(struct lustre_msg *msg);
2628 char *lustre_msg_get_jobid(struct lustre_msg *msg);
2629 __u32 lustre_msg_get_cksum(struct lustre_msg *msg);
2630 __u32 lustre_msg_calc_cksum(struct lustre_msg *msg);
2631 void lustre_msg_set_handle(struct lustre_msg *msg,
2632 struct lustre_handle *handle);
2633 void lustre_msg_set_type(struct lustre_msg *msg, __u32 type);
2634 void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc);
2635 void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid);
2636 void lustre_msg_set_last_committed(struct lustre_msg *msg,
2637 __u64 last_committed);
2638 void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions);
2639 void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno);
2640 void lustre_msg_set_status(struct lustre_msg *msg, __u32 status);
2641 void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt);
2642 void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes);
2643 void ptlrpc_request_set_replen(struct ptlrpc_request *req);
2644 void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout);
2645 void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time);
2646 void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid);
2647 void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum);
2650 lustre_shrink_reply(struct ptlrpc_request *req, int segment,
2651 unsigned int newlen, int move_data)
2653 LASSERT(req->rq_reply_state);
2654 LASSERT(req->rq_repmsg);
2655 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment,
2659 #ifdef CONFIG_LUSTRE_TRANSLATE_ERRNOS
2661 static inline int ptlrpc_status_hton(int h)
2664 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2665 * ELDLM_LOCK_ABORTED, etc.
2668 return -lustre_errno_hton(-h);
2673 static inline int ptlrpc_status_ntoh(int n)
2676 * See the comment in ptlrpc_status_hton().
2679 return -lustre_errno_ntoh(-n);
2686 #define ptlrpc_status_hton(h) (h)
2687 #define ptlrpc_status_ntoh(n) (n)
2692 /** Change request phase of \a req to \a new_phase */
2694 ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase)
2696 if (req->rq_phase == new_phase)
2699 if (new_phase == RQ_PHASE_UNREGISTERING) {
2700 req->rq_next_phase = req->rq_phase;
2702 atomic_inc(&req->rq_import->imp_unregistering);
2705 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
2707 atomic_dec(&req->rq_import->imp_unregistering);
2710 DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"",
2711 ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase));
2713 req->rq_phase = new_phase;
2717 * Returns true if request \a req got early reply and hard deadline is not met
2720 ptlrpc_client_early(struct ptlrpc_request *req)
2722 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2723 req->rq_reply_deadline > get_seconds())
2725 return req->rq_early;
2729 * Returns true if we got real reply from server for this request
2732 ptlrpc_client_replied(struct ptlrpc_request *req)
2734 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2735 req->rq_reply_deadline > get_seconds())
2737 return req->rq_replied;
2740 /** Returns true if request \a req is in process of receiving server reply */
2742 ptlrpc_client_recv(struct ptlrpc_request *req)
2744 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2745 req->rq_reply_deadline > get_seconds())
2747 return req->rq_receiving_reply;
2751 ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req)
2755 spin_lock(&req->rq_lock);
2756 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2757 req->rq_reply_deadline > get_seconds()) {
2758 spin_unlock(&req->rq_lock);
2761 rc = req->rq_receiving_reply;
2762 rc = rc || req->rq_req_unlink || req->rq_reply_unlink;
2763 spin_unlock(&req->rq_lock);
2768 ptlrpc_client_wake_req(struct ptlrpc_request *req)
2770 if (req->rq_set == NULL)
2771 wake_up(&req->rq_reply_waitq);
2773 wake_up(&req->rq_set->set_waitq);
2777 ptlrpc_rs_addref(struct ptlrpc_reply_state *rs)
2779 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2780 atomic_inc(&rs->rs_refcount);
2784 ptlrpc_rs_decref(struct ptlrpc_reply_state *rs)
2786 LASSERT(atomic_read(&rs->rs_refcount) > 0);
2787 if (atomic_dec_and_test(&rs->rs_refcount))
2788 lustre_free_reply_state(rs);
2791 /* Should only be called once per req */
2792 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req)
2794 if (req->rq_reply_state == NULL)
2795 return; /* shouldn't occur */
2796 ptlrpc_rs_decref(req->rq_reply_state);
2797 req->rq_reply_state = NULL;
2798 req->rq_repmsg = NULL;
2801 static inline __u32 lustre_request_magic(struct ptlrpc_request *req)
2803 return lustre_msg_get_magic(req->rq_reqmsg);
2806 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req)
2808 switch (req->rq_reqmsg->lm_magic) {
2809 case LUSTRE_MSG_MAGIC_V2:
2810 return req->rq_reqmsg->lm_repsize;
2812 LASSERTF(0, "incorrect message magic: %08x\n",
2813 req->rq_reqmsg->lm_magic);
2818 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req)
2820 if (req->rq_delay_limit != 0 &&
2821 time_before(cfs_time_add(req->rq_queued_time,
2822 cfs_time_seconds(req->rq_delay_limit)),
2823 cfs_time_current())) {
2829 static inline int ptlrpc_no_resend(struct ptlrpc_request *req)
2831 if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) {
2832 spin_lock(&req->rq_lock);
2833 req->rq_no_resend = 1;
2834 spin_unlock(&req->rq_lock);
2836 return req->rq_no_resend;
2840 ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt)
2842 int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate);
2844 return svcpt->scp_service->srv_watchdog_factor *
2845 max_t(int, at, obd_timeout);
2848 static inline struct ptlrpc_service *
2849 ptlrpc_req2svc(struct ptlrpc_request *req)
2851 LASSERT(req->rq_rqbd != NULL);
2852 return req->rq_rqbd->rqbd_svcpt->scp_service;
2855 /* ldlm/ldlm_lib.c */
2857 * Target client logic
2860 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg);
2861 int client_obd_cleanup(struct obd_device *obddev);
2862 int client_connect_import(const struct lu_env *env,
2863 struct obd_export **exp, struct obd_device *obd,
2864 struct obd_uuid *cluuid, struct obd_connect_data *,
2866 int client_disconnect_export(struct obd_export *exp);
2867 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
2869 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid);
2870 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
2871 struct obd_uuid *uuid);
2872 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid);
2873 void client_destroy_import(struct obd_import *imp);
2877 /* ptlrpc/pinger.c */
2879 * Pinger API (client side only)
2882 enum timeout_event {
2885 struct timeout_item;
2886 typedef int (*timeout_cb_t)(struct timeout_item *, void *);
2887 int ptlrpc_pinger_add_import(struct obd_import *imp);
2888 int ptlrpc_pinger_del_import(struct obd_import *imp);
2889 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
2890 timeout_cb_t cb, void *data,
2891 struct list_head *obd_list);
2892 int ptlrpc_del_timeout_client(struct list_head *obd_list,
2893 enum timeout_event event);
2894 struct ptlrpc_request *ptlrpc_prep_ping(struct obd_import *imp);
2895 int ptlrpc_obd_ping(struct obd_device *obd);
2896 void ping_evictor_start(void);
2897 void ping_evictor_stop(void);
2898 void ptlrpc_pinger_ir_up(void);
2899 void ptlrpc_pinger_ir_down(void);
2901 int ptlrpc_pinger_suppress_pings(void);
2903 /* ptlrpc daemon bind policy */
2905 /* all ptlrpcd threads are free mode */
2906 PDB_POLICY_NONE = 1,
2907 /* all ptlrpcd threads are bound mode */
2908 PDB_POLICY_FULL = 2,
2909 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
2910 PDB_POLICY_PAIR = 3,
2911 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
2912 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
2913 * If kernel supports NUMA, pthrpcd threads are binded and
2914 * grouped by NUMA node */
2915 PDB_POLICY_NEIGHBOR = 4,
2918 /* ptlrpc daemon load policy
2919 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
2920 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
2921 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
2922 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
2923 * depends on which is scheduled firstly, to accelerate the RPC processing. */
2925 /* on the same CPU core as the caller */
2926 PDL_POLICY_SAME = 1,
2927 /* within the same CPU partition, but not the same core as the caller */
2928 PDL_POLICY_LOCAL = 2,
2929 /* round-robin on all CPU cores, but not the same core as the caller */
2930 PDL_POLICY_ROUND = 3,
2931 /* the specified CPU core is preferred, but not enforced */
2932 PDL_POLICY_PREFERRED = 4,
2935 /* ptlrpc/ptlrpcd.c */
2936 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force);
2937 void ptlrpcd_free(struct ptlrpcd_ctl *pc);
2938 void ptlrpcd_wake(struct ptlrpc_request *req);
2939 void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx);
2940 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set);
2941 int ptlrpcd_addref(void);
2942 void ptlrpcd_decref(void);
2944 /* ptlrpc/lproc_ptlrpc.c */
2946 * procfs output related functions
2949 const char *ll_opcode2str(__u32 opcode);
2950 #if defined (CONFIG_PROC_FS)
2951 void ptlrpc_lprocfs_register_obd(struct obd_device *obd);
2952 void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd);
2953 void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes);
2955 static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {}
2956 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {}
2957 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {}
2961 /* ptlrpc/llog_client.c */
2962 extern struct llog_operations llog_client_ops;