4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 /** \defgroup PtlRPC Portal RPC and networking module.
38 * PortalRPC is the layer used by rest of lustre code to achieve network
39 * communications: establish connections with corresponding export and import
40 * states, listen for a service, send and receive RPCs.
41 * PortalRPC also includes base recovery framework: packet resending and
42 * replaying, reconnections, pinger.
44 * PortalRPC utilizes LNet as its transport layer.
58 #include <linux/lustre_net.h>
60 #include <linux/libcfs/libcfs.h>
62 #include <linux/lnet/lnet.h>
63 #include <lustre/lustre_idl.h>
64 #include <lustre_ha.h>
65 #include <lustre_sec.h>
66 #include <lustre_import.h>
67 #include <lprocfs_status.h>
68 #include <lu_object.h>
69 #include <lustre_req_layout.h>
71 #include <obd_support.h>
72 #include <lustre_ver.h>
74 /* MD flags we _always_ use */
75 #define PTLRPC_MD_OPTIONS 0
78 * Max # of bulk operations in one request.
79 * In order for the client and server to properly negotiate the maximum
80 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
81 * value. The client is free to limit the actual RPC size for any bulk
82 * transfer via cl_max_pages_per_rpc to some non-power-of-two value. */
83 #define PTLRPC_BULK_OPS_BITS 2
84 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
86 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
87 * should not be used on the server at all. Otherwise, it imposes a
88 * protocol limitation on the maximum RPC size that can be used by any
89 * RPC sent to that server in the future. Instead, the server should
90 * use the negotiated per-client ocd_brw_size to determine the bulk
92 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
95 * Define maxima for bulk I/O.
97 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
98 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
99 * currently supported maximum between peers at connect via ocd_brw_size.
101 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
102 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
103 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
105 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
106 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
107 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
108 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
109 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_CACHE_SHIFT)
110 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
112 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
113 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
114 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
116 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE))
117 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_CACHE_SIZE"
119 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
120 # error "PTLRPC_MAX_BRW_SIZE too big"
122 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
123 # error "PTLRPC_MAX_BRW_PAGES too big"
126 #define PTLRPC_NTHRS_INIT 2
131 * Constants determine how memory is used to buffer incoming service requests.
133 * ?_NBUFS # buffers to allocate when growing the pool
134 * ?_BUFSIZE # bytes in a single request buffer
135 * ?_MAXREQSIZE # maximum request service will receive
137 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
138 * of ?_NBUFS is added to the pool.
140 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
141 * considered full when less than ?_MAXREQSIZE is left in them.
146 * Constants determine how threads are created for ptlrpc service.
148 * ?_NTHRS_INIT # threads to create for each service partition on
149 * initializing. If it's non-affinity service and
150 * there is only one partition, it's the overall #
151 * threads for the service while initializing.
152 * ?_NTHRS_BASE # threads should be created at least for each
153 * ptlrpc partition to keep the service healthy.
154 * It's the low-water mark of threads upper-limit
155 * for each partition.
156 * ?_THR_FACTOR # threads can be added on threads upper-limit for
157 * each CPU core. This factor is only for reference,
158 * we might decrease value of factor if number of cores
159 * per CPT is above a limit.
160 * ?_NTHRS_MAX # overall threads can be created for a service,
161 * it's a soft limit because if service is running
162 * on machine with hundreds of cores and tens of
163 * CPU partitions, we need to guarantee each partition
164 * has ?_NTHRS_BASE threads, which means total threads
165 * will be ?_NTHRS_BASE * number_of_cpts which can
166 * exceed ?_NTHRS_MAX.
170 * #define MDS_NTHRS_INIT 2
171 * #define MDS_NTHRS_BASE 64
172 * #define MDS_NTHRS_FACTOR 8
173 * #define MDS_NTHRS_MAX 1024
176 * ---------------------------------------------------------------------
177 * Server(A) has 16 cores, user configured it to 4 partitions so each
178 * partition has 4 cores, then actual number of service threads on each
180 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
182 * Total number of threads for the service is:
183 * 96 * partitions(4) = 384
186 * ---------------------------------------------------------------------
187 * Server(B) has 32 cores, user configured it to 4 partitions so each
188 * partition has 8 cores, then actual number of service threads on each
190 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
192 * Total number of threads for the service is:
193 * 128 * partitions(4) = 512
196 * ---------------------------------------------------------------------
197 * Server(B) has 96 cores, user configured it to 8 partitions so each
198 * partition has 12 cores, then actual number of service threads on each
200 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
202 * Total number of threads for the service is:
203 * 160 * partitions(8) = 1280
205 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
206 * as upper limit of threads number for each partition:
207 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
210 * ---------------------------------------------------------------------
211 * Server(C) have a thousand of cores and user configured it to 32 partitions
212 * MDS_NTHRS_BASE(64) * 32 = 2048
214 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
215 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
216 * to keep service healthy, so total number of threads will just be 2048.
218 * NB: we don't suggest to choose server with that many cores because backend
219 * filesystem itself, buffer cache, or underlying network stack might
220 * have some SMP scalability issues at that large scale.
222 * If user already has a fat machine with hundreds or thousands of cores,
223 * there are two choices for configuration:
224 * a) create CPU table from subset of all CPUs and run Lustre on
226 * b) bind service threads on a few partitions, see modparameters of
227 * MDS and OSS for details
229 * NB: these calculations (and examples below) are simplified to help
230 * understanding, the real implementation is a little more complex,
231 * please see ptlrpc_server_nthreads_check() for details.
236 * LDLM threads constants:
238 * Given 8 as factor and 24 as base threads number
241 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
244 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
245 * threads for each partition and total threads number will be 112.
248 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
249 * threads for each partition to keep service healthy, so total threads
250 * number should be 24 * 8 = 192.
252 * So with these constants, threads number will be at the similar level
253 * of old versions, unless target machine has over a hundred cores
255 #define LDLM_THR_FACTOR 8
256 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
257 #define LDLM_NTHRS_BASE 24
258 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
260 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
261 #define LDLM_CLIENT_NBUFS 1
262 #define LDLM_SERVER_NBUFS 64
263 #define LDLM_BUFSIZE (8 * 1024)
264 #define LDLM_MAXREQSIZE (5 * 1024)
265 #define LDLM_MAXREPSIZE (1024)
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 encrupted (if needed) to
458 * produce reply buffer for actual sending. In simple case
459 * of no network encryption we jus 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 spin_is_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 bitmast 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 * CRR-N, Client Round Robin over NIDs
1191 * private data structure for CRR-N NRS
1193 struct nrs_crrn_net
{
1194 struct ptlrpc_nrs_resource cn_res
;
1195 cfs_binheap_t
*cn_binheap
;
1196 struct cfs_hash
*cn_cli_hash
;
1198 * Used when a new scheduling round commences, in order to synchronize
1199 * all clients with the new round number.
1203 * Determines the relevant ordering amongst request batches within a
1208 * Round Robin quantum; the maximum number of RPCs that each request
1209 * batch for each client can have in a scheduling round.
1215 * Object representing a client in CRR-N, as identified by its NID
1217 struct nrs_crrn_client
{
1218 struct ptlrpc_nrs_resource cc_res
;
1219 struct hlist_node cc_hnode
;
1222 * The round number against which this client is currently scheduling
1227 * The sequence number used for requests scheduled by this client during
1228 * the current round number.
1233 * Round Robin quantum; the maximum number of RPCs the client is allowed
1234 * to schedule in a single batch of each round.
1238 * # of pending requests for this client, on all existing rounds
1244 * CRR-N NRS request definition
1246 struct nrs_crrn_req
{
1248 * Round number for this request; shared with all other requests in the
1253 * Sequence number for this request; shared with all other requests in
1260 * CRR-N policy operations.
1264 * Read the RR quantum size of a CRR-N policy.
1266 NRS_CTL_CRRN_RD_QUANTUM
= PTLRPC_NRS_CTL_1ST_POL_SPEC
,
1268 * Write the RR quantum size of a CRR-N policy.
1270 NRS_CTL_CRRN_WR_QUANTUM
,
1278 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
1283 * Lower and upper byte offsets of a brw RPC
1285 struct nrs_orr_req_range
{
1291 * RPC types supported by the ORR/TRR policies
1294 NOS_OST_READ
= (1 << 0),
1295 NOS_OST_WRITE
= (1 << 1),
1296 NOS_OST_RW
= (NOS_OST_READ
| NOS_OST_WRITE
),
1298 * Default value for policies.
1300 NOS_DFLT
= NOS_OST_READ
1304 * As unique keys for grouping RPCs together, we use the object's OST FID for
1305 * the ORR policy, and the OST index for the TRR policy.
1307 * XXX: We waste some space for TRR policy instances by using a union, but it
1308 * allows to consolidate some of the code between ORR and TRR, and these
1309 * policies will probably eventually merge into one anyway.
1311 struct nrs_orr_key
{
1313 /** object FID for ORR */
1314 struct lu_fid ok_fid
;
1315 /** OST index for TRR */
1321 * The largest base string for unique hash/slab object names is
1322 * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT
1323 * id number, so this _should_ be more than enough for the maximum number of
1324 * CPTs on any system. If it does happen that this statement is incorrect,
1325 * nrs_orr_genobjname() will inevitably yield a non-unique name and cause
1326 * kmem_cache_create() to complain (on Linux), so the erroneous situation
1327 * will hopefully not go unnoticed.
1329 #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3)
1332 * private data structure for ORR and TRR NRS
1334 struct nrs_orr_data
{
1335 struct ptlrpc_nrs_resource od_res
;
1336 cfs_binheap_t
*od_binheap
;
1337 struct cfs_hash
*od_obj_hash
;
1338 struct kmem_cache
*od_cache
;
1340 * Used when a new scheduling round commences, in order to synchronize
1341 * all object or OST batches with the new round number.
1345 * Determines the relevant ordering amongst request batches within a
1350 * RPC types that are currently supported.
1352 enum nrs_orr_supp od_supp
;
1354 * Round Robin quantum; the maximum number of RPCs that each request
1355 * batch for each object or OST can have in a scheduling round.
1359 * Whether to use physical disk offsets or logical file offsets.
1363 * XXX: We need to provide a persistently allocated string to hold
1364 * unique object names for this policy, since in currently supported
1365 * versions of Linux by Lustre, kmem_cache_create() just sets a pointer
1366 * to the name string provided. kstrdup() is used in the version of
1367 * kmeme_cache_create() in current Linux mainline, so we may be able to
1368 * remove this in the future.
1370 char od_objname
[NRS_ORR_OBJ_NAME_MAX
];
1374 * Represents a backend-fs object or OST in the ORR and TRR policies
1377 struct nrs_orr_object
{
1378 struct ptlrpc_nrs_resource oo_res
;
1379 struct hlist_node oo_hnode
;
1381 * The round number against which requests are being scheduled for this
1386 * The sequence number used for requests scheduled for this object or
1387 * OST during the current round number.
1391 * The key of the object or OST for which this structure instance is
1394 struct nrs_orr_key oo_key
;
1397 * Round Robin quantum; the maximum number of RPCs that are allowed to
1398 * be scheduled for the object or OST in a single batch of each round.
1402 * # of pending requests for this object or OST, on all existing rounds
1408 * ORR/TRR NRS request definition
1410 struct nrs_orr_req
{
1412 * The offset range this request covers
1414 struct nrs_orr_req_range or_range
;
1416 * Round number for this request; shared with all other requests in the
1421 * Sequence number for this request; shared with all other requests in
1426 * For debugging purposes.
1428 struct nrs_orr_key or_key
;
1430 * An ORR policy instance has filled in request information while
1431 * enqueueing the request on the service partition's regular NRS head.
1433 unsigned int or_orr_set
:1;
1435 * A TRR policy instance has filled in request information while
1436 * enqueueing the request on the service partition's regular NRS head.
1438 unsigned int or_trr_set
:1;
1440 * Request offset ranges have been filled in with logical offset
1443 unsigned int or_logical_set
:1;
1445 * Request offset ranges have been filled in with physical offset
1448 unsigned int or_physical_set
:1;
1456 * Instances of this object exist embedded within ptlrpc_request; the main
1457 * purpose of this object is to hold references to the request's resources
1458 * for the lifetime of the request, and to hold properties that policies use
1459 * use for determining the request's scheduling priority.
1461 struct ptlrpc_nrs_request
{
1463 * The request's resource hierarchy.
1465 struct ptlrpc_nrs_resource
*nr_res_ptrs
[NRS_RES_MAX
];
1467 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1468 * policy that was used to enqueue the request.
1470 * \see nrs_request_enqueue()
1472 unsigned nr_res_idx
;
1473 unsigned nr_initialized
:1;
1474 unsigned nr_enqueued
:1;
1475 unsigned nr_started
:1;
1476 unsigned nr_finalized
:1;
1477 cfs_binheap_node_t nr_node
;
1480 * Policy-specific fields, used for determining a request's scheduling
1481 * priority, and other supporting functionality.
1485 * Fields for the FIFO policy
1487 struct nrs_fifo_req fifo
;
1489 * CRR-N request definition
1491 struct nrs_crrn_req crr
;
1492 /** ORR and TRR share the same request definition */
1493 struct nrs_orr_req orr
;
1496 * Externally-registering policies may want to use this to allocate
1497 * their own request properties.
1505 * Basic request prioritization operations structure.
1506 * The whole idea is centered around locks and RPCs that might affect locks.
1507 * When a lock is contended we try to give priority to RPCs that might lead
1508 * to fastest release of that lock.
1509 * Currently only implemented for OSTs only in a way that makes all
1510 * IO and truncate RPCs that are coming from a locked region where a lock is
1511 * contended a priority over other requests.
1513 struct ptlrpc_hpreq_ops
{
1515 * Check if the lock handle of the given lock is the same as
1516 * taken from the request.
1518 int (*hpreq_lock_match
)(struct ptlrpc_request
*, struct ldlm_lock
*);
1520 * Check if the request is a high priority one.
1522 int (*hpreq_check
)(struct ptlrpc_request
*);
1524 * Called after the request has been handled.
1526 void (*hpreq_fini
)(struct ptlrpc_request
*);
1530 * Represents remote procedure call.
1532 * This is a staple structure used by everybody wanting to send a request
1535 struct ptlrpc_request
{
1536 /* Request type: one of PTL_RPC_MSG_* */
1538 /** Result of request processing */
1541 * Linkage item through which this request is included into
1542 * sending/delayed lists on client and into rqbd list on server
1544 struct list_head rq_list
;
1546 * Server side list of incoming unserved requests sorted by arrival
1547 * time. Traversed from time to time to notice about to expire
1548 * requests and sent back "early replies" to clients to let them
1549 * know server is alive and well, just very busy to service their
1552 struct list_head rq_timed_list
;
1553 /** server-side history, used for debugging purposes. */
1554 struct list_head rq_history_list
;
1555 /** server-side per-export list */
1556 struct list_head rq_exp_list
;
1557 /** server-side hp handlers */
1558 struct ptlrpc_hpreq_ops
*rq_ops
;
1560 /** initial thread servicing this request */
1561 struct ptlrpc_thread
*rq_svc_thread
;
1563 /** history sequence # */
1564 __u64 rq_history_seq
;
1568 /** stub for NRS request */
1569 struct ptlrpc_nrs_request rq_nrq
;
1571 /** the index of service's srv_at_array into which request is linked */
1573 /** Lock to protect request flags and some other important bits, like
1577 /** client-side flags are serialized by rq_lock */
1578 unsigned int rq_intr
:1, rq_replied
:1, rq_err
:1,
1579 rq_timedout
:1, rq_resend
:1, rq_restart
:1,
1581 * when ->rq_replay is set, request is kept by the client even
1582 * after server commits corresponding transaction. This is
1583 * used for operations that require sequence of multiple
1584 * requests to be replayed. The only example currently is file
1585 * open/close. When last request in such a sequence is
1586 * committed, ->rq_replay is cleared on all requests in the
1590 rq_no_resend
:1, rq_waiting
:1, rq_receiving_reply
:1,
1591 rq_no_delay
:1, rq_net_err
:1, rq_wait_ctx
:1,
1592 rq_early
:1, rq_must_unlink
:1,
1593 rq_memalloc
:1, /* req originated from "kswapd" */
1594 /* server-side flags */
1595 rq_packed_final
:1, /* packed final reply */
1596 rq_hp
:1, /* high priority RPC */
1597 rq_at_linked
:1, /* link into service's srv_at_array */
1598 rq_reply_truncate
:1,
1600 /* whether the "rq_set" is a valid one */
1602 rq_generation_set
:1,
1603 /* do not resend request on -EINPROGRESS */
1604 rq_no_retry_einprogress
:1,
1605 /* allow the req to be sent if the import is in recovery
1609 unsigned int rq_nr_resend
;
1611 enum rq_phase rq_phase
; /* one of RQ_PHASE_* */
1612 enum rq_phase rq_next_phase
; /* one of RQ_PHASE_* to be used next */
1613 atomic_t rq_refcount
;/* client-side refcount for SENT race,
1614 server-side refcount for multiple replies */
1616 /** Portal to which this request would be sent */
1617 short rq_request_portal
; /* XXX FIXME bug 249 */
1618 /** Portal where to wait for reply and where reply would be sent */
1619 short rq_reply_portal
; /* XXX FIXME bug 249 */
1623 * !rq_truncate : # reply bytes actually received,
1624 * rq_truncate : required repbuf_len for resend
1626 int rq_nob_received
;
1627 /** Request length */
1631 /** Request message - what client sent */
1632 struct lustre_msg
*rq_reqmsg
;
1633 /** Reply message - server response */
1634 struct lustre_msg
*rq_repmsg
;
1635 /** Transaction number */
1640 * List item to for replay list. Not yet committed requests get linked
1642 * Also see \a rq_replay comment above.
1644 struct list_head rq_replay_list
;
1647 * security and encryption data
1649 struct ptlrpc_cli_ctx
*rq_cli_ctx
; /**< client's half ctx */
1650 struct ptlrpc_svc_ctx
*rq_svc_ctx
; /**< server's half ctx */
1651 struct list_head rq_ctx_chain
; /**< link to waited ctx */
1653 struct sptlrpc_flavor rq_flvr
; /**< for client & server */
1654 enum lustre_sec_part rq_sp_from
;
1656 /* client/server security flags */
1658 rq_ctx_init
:1, /* context initiation */
1659 rq_ctx_fini
:1, /* context destroy */
1660 rq_bulk_read
:1, /* request bulk read */
1661 rq_bulk_write
:1, /* request bulk write */
1662 /* server authentication flags */
1663 rq_auth_gss
:1, /* authenticated by gss */
1664 rq_auth_remote
:1, /* authed as remote user */
1665 rq_auth_usr_root
:1, /* authed as root */
1666 rq_auth_usr_mdt
:1, /* authed as mdt */
1667 rq_auth_usr_ost
:1, /* authed as ost */
1668 /* security tfm flags */
1671 /* doesn't expect reply FIXME */
1673 rq_pill_init
:1; /* pill initialized */
1675 uid_t rq_auth_uid
; /* authed uid */
1676 uid_t rq_auth_mapped_uid
; /* authed uid mapped to */
1678 /* (server side), pointed directly into req buffer */
1679 struct ptlrpc_user_desc
*rq_user_desc
;
1681 /* various buffer pointers */
1682 struct lustre_msg
*rq_reqbuf
; /* req wrapper */
1683 char *rq_repbuf
; /* rep buffer */
1684 struct lustre_msg
*rq_repdata
; /* rep wrapper msg */
1685 struct lustre_msg
*rq_clrbuf
; /* only in priv mode */
1686 int rq_reqbuf_len
; /* req wrapper buf len */
1687 int rq_reqdata_len
; /* req wrapper msg len */
1688 int rq_repbuf_len
; /* rep buffer len */
1689 int rq_repdata_len
; /* rep wrapper msg len */
1690 int rq_clrbuf_len
; /* only in priv mode */
1691 int rq_clrdata_len
; /* only in priv mode */
1693 /** early replies go to offset 0, regular replies go after that */
1694 unsigned int rq_reply_off
;
1698 /** Fields that help to see if request and reply were swabbed or not */
1699 __u32 rq_req_swab_mask
;
1700 __u32 rq_rep_swab_mask
;
1702 /** What was import generation when this request was sent */
1703 int rq_import_generation
;
1704 enum lustre_imp_state rq_send_state
;
1706 /** how many early replies (for stats) */
1709 /** client+server request */
1710 lnet_handle_md_t rq_req_md_h
;
1711 struct ptlrpc_cb_id rq_req_cbid
;
1712 /** optional time limit for send attempts */
1713 cfs_duration_t rq_delay_limit
;
1714 /** time request was first queued */
1715 cfs_time_t rq_queued_time
;
1717 /* server-side... */
1718 /** request arrival time */
1719 struct timeval rq_arrival_time
;
1720 /** separated reply state */
1721 struct ptlrpc_reply_state
*rq_reply_state
;
1722 /** incoming request buffer */
1723 struct ptlrpc_request_buffer_desc
*rq_rqbd
;
1725 /** client-only incoming reply */
1726 lnet_handle_md_t rq_reply_md_h
;
1727 wait_queue_head_t rq_reply_waitq
;
1728 struct ptlrpc_cb_id rq_reply_cbid
;
1732 /** Peer description (the other side) */
1733 lnet_process_id_t rq_peer
;
1734 /** Server-side, export on which request was received */
1735 struct obd_export
*rq_export
;
1736 /** Client side, import where request is being sent */
1737 struct obd_import
*rq_import
;
1739 /** Replay callback, called after request is replayed at recovery */
1740 void (*rq_replay_cb
)(struct ptlrpc_request
*);
1742 * Commit callback, called when request is committed and about to be
1745 void (*rq_commit_cb
)(struct ptlrpc_request
*);
1746 /** Opaq data for replay and commit callbacks. */
1749 /** For bulk requests on client only: bulk descriptor */
1750 struct ptlrpc_bulk_desc
*rq_bulk
;
1752 /** client outgoing req */
1754 * when request/reply sent (secs), or time when request should be sent
1757 /** time for request really sent out */
1758 time_t rq_real_sent
;
1760 /** when request must finish. volatile
1761 * so that servers' early reply updates to the deadline aren't
1762 * kept in per-cpu cache */
1763 volatile time_t rq_deadline
;
1764 /** when req reply unlink must finish. */
1765 time_t rq_reply_deadline
;
1766 /** when req bulk unlink must finish. */
1767 time_t rq_bulk_deadline
;
1769 * service time estimate (secs)
1770 * If the requestsis not served by this time, it is marked as timed out.
1774 /** Multi-rpc bits */
1775 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1776 wait_queue_head_t rq_set_waitq
;
1777 /** Link item for request set lists */
1778 struct list_head rq_set_chain
;
1779 /** Link back to the request set */
1780 struct ptlrpc_request_set
*rq_set
;
1781 /** Async completion handler, called when reply is received */
1782 ptlrpc_interpterer_t rq_interpret_reply
;
1783 /** Async completion context */
1784 union ptlrpc_async_args rq_async_args
;
1786 /** Pool if request is from preallocated list */
1787 struct ptlrpc_request_pool
*rq_pool
;
1789 struct lu_context rq_session
;
1790 struct lu_context rq_recov_session
;
1792 /** request format description */
1793 struct req_capsule rq_pill
;
1797 * Call completion handler for rpc if any, return it's status or original
1798 * rc if there was no handler defined for this request.
1800 static inline int ptlrpc_req_interpret(const struct lu_env
*env
,
1801 struct ptlrpc_request
*req
, int rc
)
1803 if (req
->rq_interpret_reply
!= NULL
) {
1804 req
->rq_status
= req
->rq_interpret_reply(env
, req
,
1805 &req
->rq_async_args
,
1807 return req
->rq_status
;
1815 int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf
*conf
);
1816 int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf
*conf
);
1817 void ptlrpc_nrs_req_hp_move(struct ptlrpc_request
*req
);
1818 void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy
*policy
,
1819 struct ptlrpc_nrs_pol_info
*info
);
1822 * Can the request be moved from the regular NRS head to the high-priority NRS
1823 * head (of the same PTLRPC service partition), if any?
1825 * For a reliable result, this should be checked under svcpt->scp_req lock.
1827 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request
*req
)
1829 struct ptlrpc_nrs_request
*nrq
= &req
->rq_nrq
;
1832 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1833 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1834 * to make sure it has not been scheduled yet (analogous to previous
1835 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1837 return nrq
->nr_enqueued
&& !nrq
->nr_started
&& !req
->rq_hp
;
1842 * Returns 1 if request buffer at offset \a index was already swabbed
1844 static inline int lustre_req_swabbed(struct ptlrpc_request
*req
, int index
)
1846 LASSERT(index
< sizeof(req
->rq_req_swab_mask
) * 8);
1847 return req
->rq_req_swab_mask
& (1 << index
);
1851 * Returns 1 if request reply buffer at offset \a index was already swabbed
1853 static inline int lustre_rep_swabbed(struct ptlrpc_request
*req
, int index
)
1855 LASSERT(index
< sizeof(req
->rq_rep_swab_mask
) * 8);
1856 return req
->rq_rep_swab_mask
& (1 << index
);
1860 * Returns 1 if request needs to be swabbed into local cpu byteorder
1862 static inline int ptlrpc_req_need_swab(struct ptlrpc_request
*req
)
1864 return lustre_req_swabbed(req
, MSG_PTLRPC_HEADER_OFF
);
1868 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1870 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request
*req
)
1872 return lustre_rep_swabbed(req
, MSG_PTLRPC_HEADER_OFF
);
1876 * Mark request buffer at offset \a index that it was already swabbed
1878 static inline void lustre_set_req_swabbed(struct ptlrpc_request
*req
, int index
)
1880 LASSERT(index
< sizeof(req
->rq_req_swab_mask
) * 8);
1881 LASSERT((req
->rq_req_swab_mask
& (1 << index
)) == 0);
1882 req
->rq_req_swab_mask
|= 1 << index
;
1886 * Mark request reply buffer at offset \a index that it was already swabbed
1888 static inline void lustre_set_rep_swabbed(struct ptlrpc_request
*req
, int index
)
1890 LASSERT(index
< sizeof(req
->rq_rep_swab_mask
) * 8);
1891 LASSERT((req
->rq_rep_swab_mask
& (1 << index
)) == 0);
1892 req
->rq_rep_swab_mask
|= 1 << index
;
1896 * Convert numerical request phase value \a phase into text string description
1898 static inline const char *
1899 ptlrpc_phase2str(enum rq_phase phase
)
1908 case RQ_PHASE_INTERPRET
:
1910 case RQ_PHASE_COMPLETE
:
1912 case RQ_PHASE_UNREGISTERING
:
1913 return "Unregistering";
1920 * Convert numerical request phase of the request \a req into text stringi
1923 static inline const char *
1924 ptlrpc_rqphase2str(struct ptlrpc_request
*req
)
1926 return ptlrpc_phase2str(req
->rq_phase
);
1930 * Debugging functions and helpers to print request structure into debug log
1933 /* Spare the preprocessor, spoil the bugs. */
1934 #define FLAG(field, str) (field ? str : "")
1936 /** Convert bit flags into a string */
1937 #define DEBUG_REQ_FLAGS(req) \
1938 ptlrpc_rqphase2str(req), \
1939 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1940 FLAG(req->rq_err, "E"), \
1941 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1942 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1943 FLAG(req->rq_no_resend, "N"), \
1944 FLAG(req->rq_waiting, "W"), \
1945 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1946 FLAG(req->rq_committed, "M")
1948 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1950 void _debug_req(struct ptlrpc_request
*req
,
1951 struct libcfs_debug_msg_data
*data
, const char *fmt
, ...)
1952 __attribute__ ((format (printf
, 3, 4)));
1955 * Helper that decides if we need to print request according to current debug
1958 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1960 CFS_CHECK_STACK(msgdata, mask, cdls); \
1962 if (((mask) & D_CANTMASK) != 0 || \
1963 ((libcfs_debug & (mask)) != 0 && \
1964 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1965 _debug_req((req), msgdata, fmt, ##a); \
1969 * This is the debug print function you need to use to print request structure
1970 * content into lustre debug log.
1971 * for most callers (level is a constant) this is resolved at compile time */
1972 #define DEBUG_REQ(level, req, fmt, args...) \
1974 if ((level) & (D_ERROR | D_WARNING)) { \
1975 static struct cfs_debug_limit_state cdls; \
1976 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1977 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1979 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1980 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1986 * Structure that defines a single page of a bulk transfer
1988 struct ptlrpc_bulk_page
{
1989 /** Linkage to list of pages in a bulk */
1990 struct list_head bp_link
;
1992 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1995 /** offset within a page */
1997 /** The page itself */
1998 struct page
*bp_page
;
2001 #define BULK_GET_SOURCE 0
2002 #define BULK_PUT_SINK 1
2003 #define BULK_GET_SINK 2
2004 #define BULK_PUT_SOURCE 3
2007 * Definition of bulk descriptor.
2008 * Bulks are special "Two phase" RPCs where initial request message
2009 * is sent first and it is followed bt a transfer (o receiving) of a large
2010 * amount of data to be settled into pages referenced from the bulk descriptors.
2011 * Bulks transfers (the actual data following the small requests) are done
2012 * on separate LNet portals.
2013 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
2014 * Another user is readpage for MDT.
2016 struct ptlrpc_bulk_desc
{
2017 /** completed with failure */
2018 unsigned long bd_failure
:1;
2019 /** {put,get}{source,sink} */
2020 unsigned long bd_type
:2;
2022 unsigned long bd_registered
:1;
2023 /** For serialization with callback */
2025 /** Import generation when request for this bulk was sent */
2026 int bd_import_generation
;
2027 /** LNet portal for this bulk */
2029 /** Server side - export this bulk created for */
2030 struct obd_export
*bd_export
;
2031 /** Client side - import this bulk was sent on */
2032 struct obd_import
*bd_import
;
2033 /** Back pointer to the request */
2034 struct ptlrpc_request
*bd_req
;
2035 wait_queue_head_t bd_waitq
; /* server side only WQ */
2036 int bd_iov_count
; /* # entries in bd_iov */
2037 int bd_max_iov
; /* allocated size of bd_iov */
2038 int bd_nob
; /* # bytes covered */
2039 int bd_nob_transferred
; /* # bytes GOT/PUT */
2043 struct ptlrpc_cb_id bd_cbid
; /* network callback info */
2044 lnet_nid_t bd_sender
; /* stash event::sender */
2045 int bd_md_count
; /* # valid entries in bd_mds */
2046 int bd_md_max_brw
; /* max entries in bd_mds */
2047 /** array of associated MDs */
2048 lnet_handle_md_t bd_mds
[PTLRPC_BULK_OPS_COUNT
];
2051 * encrypt iov, size is either 0 or bd_iov_count.
2053 lnet_kiov_t
*bd_enc_iov
;
2055 lnet_kiov_t bd_iov
[0];
2059 SVC_STOPPED
= 1 << 0,
2060 SVC_STOPPING
= 1 << 1,
2061 SVC_STARTING
= 1 << 2,
2062 SVC_RUNNING
= 1 << 3,
2064 SVC_SIGNAL
= 1 << 5,
2067 #define PTLRPC_THR_NAME_LEN 32
2069 * Definition of server service thread structure
2071 struct ptlrpc_thread
{
2073 * List of active threads in svc->srv_threads
2075 struct list_head t_link
;
2077 * thread-private data (preallocated memory)
2082 * service thread index, from ptlrpc_start_threads
2086 * service thread pid
2090 * put watchdog in the structure per thread b=14840
2092 * Lustre watchdog is removed for client in the hope
2093 * of a generic watchdog can be merged in kernel.
2094 * When that happens, we should add below back.
2096 * struct lc_watchdog *t_watchdog;
2099 * the svc this thread belonged to b=18582
2101 struct ptlrpc_service_part
*t_svcpt
;
2102 wait_queue_head_t t_ctl_waitq
;
2103 struct lu_env
*t_env
;
2104 char t_name
[PTLRPC_THR_NAME_LEN
];
2107 static inline int thread_is_init(struct ptlrpc_thread
*thread
)
2109 return thread
->t_flags
== 0;
2112 static inline int thread_is_stopped(struct ptlrpc_thread
*thread
)
2114 return !!(thread
->t_flags
& SVC_STOPPED
);
2117 static inline int thread_is_stopping(struct ptlrpc_thread
*thread
)
2119 return !!(thread
->t_flags
& SVC_STOPPING
);
2122 static inline int thread_is_starting(struct ptlrpc_thread
*thread
)
2124 return !!(thread
->t_flags
& SVC_STARTING
);
2127 static inline int thread_is_running(struct ptlrpc_thread
*thread
)
2129 return !!(thread
->t_flags
& SVC_RUNNING
);
2132 static inline int thread_is_event(struct ptlrpc_thread
*thread
)
2134 return !!(thread
->t_flags
& SVC_EVENT
);
2137 static inline int thread_is_signal(struct ptlrpc_thread
*thread
)
2139 return !!(thread
->t_flags
& SVC_SIGNAL
);
2142 static inline void thread_clear_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
2144 thread
->t_flags
&= ~flags
;
2147 static inline void thread_set_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
2149 thread
->t_flags
= flags
;
2152 static inline void thread_add_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
2154 thread
->t_flags
|= flags
;
2157 static inline int thread_test_and_clear_flags(struct ptlrpc_thread
*thread
,
2160 if (thread
->t_flags
& flags
) {
2161 thread
->t_flags
&= ~flags
;
2168 * Request buffer descriptor structure.
2169 * This is a structure that contains one posted request buffer for service.
2170 * Once data land into a buffer, event callback creates actual request and
2171 * notifies wakes one of the service threads to process new incoming request.
2172 * More than one request can fit into the buffer.
2174 struct ptlrpc_request_buffer_desc
{
2175 /** Link item for rqbds on a service */
2176 struct list_head rqbd_list
;
2177 /** History of requests for this buffer */
2178 struct list_head rqbd_reqs
;
2179 /** Back pointer to service for which this buffer is registered */
2180 struct ptlrpc_service_part
*rqbd_svcpt
;
2181 /** LNet descriptor */
2182 lnet_handle_md_t rqbd_md_h
;
2184 /** The buffer itself */
2186 struct ptlrpc_cb_id rqbd_cbid
;
2188 * This "embedded" request structure is only used for the
2189 * last request to fit into the buffer
2191 struct ptlrpc_request rqbd_req
;
2194 typedef int (*svc_handler_t
)(struct ptlrpc_request
*req
);
2196 struct ptlrpc_service_ops
{
2198 * if non-NULL called during thread creation (ptlrpc_start_thread())
2199 * to initialize service specific per-thread state.
2201 int (*so_thr_init
)(struct ptlrpc_thread
*thr
);
2203 * if non-NULL called during thread shutdown (ptlrpc_main()) to
2204 * destruct state created by ->srv_init().
2206 void (*so_thr_done
)(struct ptlrpc_thread
*thr
);
2208 * Handler function for incoming requests for this service
2210 int (*so_req_handler
)(struct ptlrpc_request
*req
);
2212 * function to determine priority of the request, it's called
2213 * on every new request
2215 int (*so_hpreq_handler
)(struct ptlrpc_request
*);
2217 * service-specific print fn
2219 void (*so_req_printer
)(void *, struct ptlrpc_request
*);
2222 #ifndef __cfs_cacheline_aligned
2223 /* NB: put it here for reducing patche dependence */
2224 # define __cfs_cacheline_aligned
2228 * How many high priority requests to serve before serving one normal
2231 #define PTLRPC_SVC_HP_RATIO 10
2234 * Definition of PortalRPC service.
2235 * The service is listening on a particular portal (like tcp port)
2236 * and perform actions for a specific server like IO service for OST
2237 * or general metadata service for MDS.
2239 struct ptlrpc_service
{
2240 /** serialize /proc operations */
2241 spinlock_t srv_lock
;
2242 /** most often accessed fields */
2243 /** chain thru all services */
2244 struct list_head srv_list
;
2245 /** service operations table */
2246 struct ptlrpc_service_ops srv_ops
;
2247 /** only statically allocated strings here; we don't clean them */
2249 /** only statically allocated strings here; we don't clean them */
2250 char *srv_thread_name
;
2251 /** service thread list */
2252 struct list_head srv_threads
;
2253 /** threads # should be created for each partition on initializing */
2254 int srv_nthrs_cpt_init
;
2255 /** limit of threads number for each partition */
2256 int srv_nthrs_cpt_limit
;
2257 /** Root of /proc dir tree for this service */
2258 struct proc_dir_entry
*srv_procroot
;
2259 /** Pointer to statistic data for this service */
2260 struct lprocfs_stats
*srv_stats
;
2261 /** # hp per lp reqs to handle */
2262 int srv_hpreq_ratio
;
2263 /** biggest request to receive */
2264 int srv_max_req_size
;
2265 /** biggest reply to send */
2266 int srv_max_reply_size
;
2267 /** size of individual buffers */
2269 /** # buffers to allocate in 1 group */
2270 int srv_nbuf_per_group
;
2271 /** Local portal on which to receive requests */
2272 __u32 srv_req_portal
;
2273 /** Portal on the client to send replies to */
2274 __u32 srv_rep_portal
;
2276 * Tags for lu_context associated with this thread, see struct
2280 /** soft watchdog timeout multiplier */
2281 int srv_watchdog_factor
;
2282 /** under unregister_service */
2283 unsigned srv_is_stopping
:1;
2285 /** max # request buffers in history per partition */
2286 int srv_hist_nrqbds_cpt_max
;
2287 /** number of CPTs this service bound on */
2289 /** CPTs array this service bound on */
2291 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2293 /** CPT table this service is running over */
2294 struct cfs_cpt_table
*srv_cptable
;
2296 * partition data for ptlrpc service
2298 struct ptlrpc_service_part
*srv_parts
[0];
2302 * Definition of PortalRPC service partition data.
2303 * Although a service only has one instance of it right now, but we
2304 * will have multiple instances very soon (instance per CPT).
2306 * it has four locks:
2308 * serialize operations on rqbd and requests waiting for preprocess
2310 * serialize operations active requests sent to this portal
2312 * serialize adaptive timeout stuff
2314 * serialize operations on RS list (reply states)
2316 * We don't have any use-case to take two or more locks at the same time
2317 * for now, so there is no lock order issue.
2319 struct ptlrpc_service_part
{
2320 /** back reference to owner */
2321 struct ptlrpc_service
*scp_service __cfs_cacheline_aligned
;
2322 /* CPT id, reserved */
2324 /** always increasing number */
2326 /** # of starting threads */
2327 int scp_nthrs_starting
;
2328 /** # of stopping threads, reserved for shrinking threads */
2329 int scp_nthrs_stopping
;
2330 /** # running threads */
2331 int scp_nthrs_running
;
2332 /** service threads list */
2333 struct list_head scp_threads
;
2336 * serialize the following fields, used for protecting
2337 * rqbd list and incoming requests waiting for preprocess,
2338 * threads starting & stopping are also protected by this lock.
2340 spinlock_t scp_lock __cfs_cacheline_aligned
;
2341 /** total # req buffer descs allocated */
2342 int scp_nrqbds_total
;
2343 /** # posted request buffers for receiving */
2344 int scp_nrqbds_posted
;
2345 /** in progress of allocating rqbd */
2346 int scp_rqbd_allocating
;
2347 /** # incoming reqs */
2348 int scp_nreqs_incoming
;
2349 /** request buffers to be reposted */
2350 struct list_head scp_rqbd_idle
;
2351 /** req buffers receiving */
2352 struct list_head scp_rqbd_posted
;
2353 /** incoming reqs */
2354 struct list_head scp_req_incoming
;
2355 /** timeout before re-posting reqs, in tick */
2356 cfs_duration_t scp_rqbd_timeout
;
2358 * all threads sleep on this. This wait-queue is signalled when new
2359 * incoming request arrives and when difficult reply has to be handled.
2361 wait_queue_head_t scp_waitq
;
2363 /** request history */
2364 struct list_head scp_hist_reqs
;
2365 /** request buffer history */
2366 struct list_head scp_hist_rqbds
;
2367 /** # request buffers in history */
2368 int scp_hist_nrqbds
;
2369 /** sequence number for request */
2371 /** highest seq culled from history */
2372 __u64 scp_hist_seq_culled
;
2375 * serialize the following fields, used for processing requests
2376 * sent to this portal
2378 spinlock_t scp_req_lock __cfs_cacheline_aligned
;
2379 /** # reqs in either of the NRS heads below */
2380 /** # reqs being served */
2381 int scp_nreqs_active
;
2382 /** # HPreqs being served */
2383 int scp_nhreqs_active
;
2384 /** # hp requests handled */
2387 /** NRS head for regular requests */
2388 struct ptlrpc_nrs scp_nrs_reg
;
2389 /** NRS head for HP requests; this is only valid for services that can
2390 * handle HP requests */
2391 struct ptlrpc_nrs
*scp_nrs_hp
;
2396 * serialize the following fields, used for changes on
2399 spinlock_t scp_at_lock __cfs_cacheline_aligned
;
2400 /** estimated rpc service time */
2401 struct adaptive_timeout scp_at_estimate
;
2402 /** reqs waiting for replies */
2403 struct ptlrpc_at_array scp_at_array
;
2404 /** early reply timer */
2405 struct timer_list scp_at_timer
;
2407 cfs_time_t scp_at_checktime
;
2408 /** check early replies */
2409 unsigned scp_at_check
;
2413 * serialize the following fields, used for processing
2414 * replies for this portal
2416 spinlock_t scp_rep_lock __cfs_cacheline_aligned
;
2417 /** all the active replies */
2418 struct list_head scp_rep_active
;
2419 /** List of free reply_states */
2420 struct list_head scp_rep_idle
;
2421 /** waitq to run, when adding stuff to srv_free_rs_list */
2422 wait_queue_head_t scp_rep_waitq
;
2423 /** # 'difficult' replies */
2424 atomic_t scp_nreps_difficult
;
2427 #define ptlrpc_service_for_each_part(part, i, svc) \
2429 i < (svc)->srv_ncpts && \
2430 (svc)->srv_parts != NULL && \
2431 ((part) = (svc)->srv_parts[i]) != NULL; i++)
2434 * Declaration of ptlrpcd control structure
2436 struct ptlrpcd_ctl
{
2438 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2440 unsigned long pc_flags
;
2442 * Thread lock protecting structure fields.
2448 struct completion pc_starting
;
2452 struct completion pc_finishing
;
2454 * Thread requests set.
2456 struct ptlrpc_request_set
*pc_set
;
2458 * Thread name used in cfs_daemonize()
2462 * Environment for request interpreters to run in.
2464 struct lu_env pc_env
;
2466 * Index of ptlrpcd thread in the array.
2470 * Number of the ptlrpcd's partners.
2474 * Pointer to the array of partners' ptlrpcd_ctl structure.
2476 struct ptlrpcd_ctl
**pc_partners
;
2478 * Record the partner index to be processed next.
2483 /* Bits for pc_flags */
2484 enum ptlrpcd_ctl_flags
{
2486 * Ptlrpc thread start flag.
2488 LIOD_START
= 1 << 0,
2490 * Ptlrpc thread stop flag.
2494 * Ptlrpc thread force flag (only stop force so far).
2495 * This will cause aborting any inflight rpcs handled
2496 * by thread if LIOD_STOP is specified.
2498 LIOD_FORCE
= 1 << 2,
2500 * This is a recovery ptlrpc thread.
2502 LIOD_RECOVERY
= 1 << 3,
2504 * The ptlrpcd is bound to some CPU core.
2513 * Service compatibility function; the policy is compatible with all services.
2515 * \param[in] svc The service the policy is attempting to register with.
2516 * \param[in] desc The policy descriptor
2518 * \retval true The policy is compatible with the service
2520 * \see ptlrpc_nrs_pol_desc::pd_compat()
2522 static inline bool nrs_policy_compat_all(const struct ptlrpc_service
*svc
,
2523 const struct ptlrpc_nrs_pol_desc
*desc
)
2529 * Service compatibility function; the policy is compatible with only a specific
2530 * service which is identified by its human-readable name at
2531 * ptlrpc_service::srv_name.
2533 * \param[in] svc The service the policy is attempting to register with.
2534 * \param[in] desc The policy descriptor
2536 * \retval false The policy is not compatible with the service
2537 * \retval true The policy is compatible with the service
2539 * \see ptlrpc_nrs_pol_desc::pd_compat()
2541 static inline bool nrs_policy_compat_one(const struct ptlrpc_service
*svc
,
2542 const struct ptlrpc_nrs_pol_desc
*desc
)
2544 LASSERT(desc
->pd_compat_svc_name
!= NULL
);
2545 return strcmp(svc
->srv_name
, desc
->pd_compat_svc_name
) == 0;
2550 /* ptlrpc/events.c */
2551 extern lnet_handle_eq_t ptlrpc_eq_h
;
2552 extern int ptlrpc_uuid_to_peer(struct obd_uuid
*uuid
,
2553 lnet_process_id_t
*peer
, lnet_nid_t
*self
);
2555 * These callbacks are invoked by LNet when something happened to
2559 extern void request_out_callback(lnet_event_t
*ev
);
2560 extern void reply_in_callback(lnet_event_t
*ev
);
2561 extern void client_bulk_callback(lnet_event_t
*ev
);
2562 extern void request_in_callback(lnet_event_t
*ev
);
2563 extern void reply_out_callback(lnet_event_t
*ev
);
2566 /* ptlrpc/connection.c */
2567 struct ptlrpc_connection
*ptlrpc_connection_get(lnet_process_id_t peer
,
2569 struct obd_uuid
*uuid
);
2570 int ptlrpc_connection_put(struct ptlrpc_connection
*c
);
2571 struct ptlrpc_connection
*ptlrpc_connection_addref(struct ptlrpc_connection
*);
2572 int ptlrpc_connection_init(void);
2573 void ptlrpc_connection_fini(void);
2574 extern lnet_pid_t
ptl_get_pid(void);
2576 /* ptlrpc/niobuf.c */
2578 * Actual interfacing with LNet to put/get/register/unregister stuff
2582 int ptlrpc_register_bulk(struct ptlrpc_request
*req
);
2583 int ptlrpc_unregister_bulk(struct ptlrpc_request
*req
, int async
);
2585 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request
*req
)
2587 struct ptlrpc_bulk_desc
*desc
;
2590 LASSERT(req
!= NULL
);
2591 desc
= req
->rq_bulk
;
2593 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
) &&
2594 req
->rq_bulk_deadline
> cfs_time_current_sec())
2600 spin_lock(&desc
->bd_lock
);
2601 rc
= desc
->bd_md_count
;
2602 spin_unlock(&desc
->bd_lock
);
2606 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2607 #define PTLRPC_REPLY_EARLY 0x02
2608 int ptlrpc_send_reply(struct ptlrpc_request
*req
, int flags
);
2609 int ptlrpc_reply(struct ptlrpc_request
*req
);
2610 int ptlrpc_send_error(struct ptlrpc_request
*req
, int difficult
);
2611 int ptlrpc_error(struct ptlrpc_request
*req
);
2612 void ptlrpc_resend_req(struct ptlrpc_request
*request
);
2613 int ptlrpc_at_get_net_latency(struct ptlrpc_request
*req
);
2614 int ptl_send_rpc(struct ptlrpc_request
*request
, int noreply
);
2615 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc
*rqbd
);
2618 /* ptlrpc/client.c */
2620 * Client-side portals API. Everything to send requests, receive replies,
2621 * request queues, request management, etc.
2624 void ptlrpc_request_committed(struct ptlrpc_request
*req
, int force
);
2626 void ptlrpc_init_client(int req_portal
, int rep_portal
, char *name
,
2627 struct ptlrpc_client
*);
2628 void ptlrpc_cleanup_client(struct obd_import
*imp
);
2629 struct ptlrpc_connection
*ptlrpc_uuid_to_connection(struct obd_uuid
*uuid
);
2631 int ptlrpc_queue_wait(struct ptlrpc_request
*req
);
2632 int ptlrpc_replay_req(struct ptlrpc_request
*req
);
2633 int ptlrpc_unregister_reply(struct ptlrpc_request
*req
, int async
);
2634 void ptlrpc_restart_req(struct ptlrpc_request
*req
);
2635 void ptlrpc_abort_inflight(struct obd_import
*imp
);
2636 void ptlrpc_cleanup_imp(struct obd_import
*imp
);
2637 void ptlrpc_abort_set(struct ptlrpc_request_set
*set
);
2639 struct ptlrpc_request_set
*ptlrpc_prep_set(void);
2640 struct ptlrpc_request_set
*ptlrpc_prep_fcset(int max
, set_producer_func func
,
2642 int ptlrpc_set_add_cb(struct ptlrpc_request_set
*set
,
2643 set_interpreter_func fn
, void *data
);
2644 int ptlrpc_set_next_timeout(struct ptlrpc_request_set
*);
2645 int ptlrpc_check_set(const struct lu_env
*env
, struct ptlrpc_request_set
*set
);
2646 int ptlrpc_set_wait(struct ptlrpc_request_set
*);
2647 int ptlrpc_expired_set(void *data
);
2648 void ptlrpc_interrupted_set(void *data
);
2649 void ptlrpc_mark_interrupted(struct ptlrpc_request
*req
);
2650 void ptlrpc_set_destroy(struct ptlrpc_request_set
*);
2651 void ptlrpc_set_add_req(struct ptlrpc_request_set
*, struct ptlrpc_request
*);
2652 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl
*pc
,
2653 struct ptlrpc_request
*req
);
2655 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool
*pool
);
2656 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool
*pool
, int num_rq
);
2658 struct ptlrpc_request_pool
*
2659 ptlrpc_init_rq_pool(int, int,
2660 void (*populate_pool
)(struct ptlrpc_request_pool
*, int));
2662 void ptlrpc_at_set_req_timeout(struct ptlrpc_request
*req
);
2663 struct ptlrpc_request
*ptlrpc_request_alloc(struct obd_import
*imp
,
2664 const struct req_format
*format
);
2665 struct ptlrpc_request
*ptlrpc_request_alloc_pool(struct obd_import
*imp
,
2666 struct ptlrpc_request_pool
*,
2667 const struct req_format
*format
);
2668 void ptlrpc_request_free(struct ptlrpc_request
*request
);
2669 int ptlrpc_request_pack(struct ptlrpc_request
*request
,
2670 __u32 version
, int opcode
);
2671 struct ptlrpc_request
*ptlrpc_request_alloc_pack(struct obd_import
*imp
,
2672 const struct req_format
*format
,
2673 __u32 version
, int opcode
);
2674 int ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
2675 __u32 version
, int opcode
, char **bufs
,
2676 struct ptlrpc_cli_ctx
*ctx
);
2677 struct ptlrpc_request
*ptlrpc_prep_req(struct obd_import
*imp
, __u32 version
,
2678 int opcode
, int count
, __u32
*lengths
,
2680 struct ptlrpc_request
*ptlrpc_prep_req_pool(struct obd_import
*imp
,
2681 __u32 version
, int opcode
,
2682 int count
, __u32
*lengths
, char **bufs
,
2683 struct ptlrpc_request_pool
*pool
);
2684 void ptlrpc_req_finished(struct ptlrpc_request
*request
);
2685 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request
*request
);
2686 struct ptlrpc_request
*ptlrpc_request_addref(struct ptlrpc_request
*req
);
2687 struct ptlrpc_bulk_desc
*ptlrpc_prep_bulk_imp(struct ptlrpc_request
*req
,
2688 unsigned npages
, unsigned max_brw
,
2689 unsigned type
, unsigned portal
);
2690 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc
*bulk
, int pin
);
2691 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc
*bulk
)
2693 __ptlrpc_free_bulk(bulk
, 1);
2695 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc
*bulk
)
2697 __ptlrpc_free_bulk(bulk
, 0);
2699 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc
*desc
,
2700 struct page
*page
, int pageoffset
, int len
, int);
2701 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc
*desc
,
2702 struct page
*page
, int pageoffset
,
2705 __ptlrpc_prep_bulk_page(desc
, page
, pageoffset
, len
, 1);
2708 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc
*desc
,
2709 struct page
*page
, int pageoffset
,
2712 __ptlrpc_prep_bulk_page(desc
, page
, pageoffset
, len
, 0);
2715 void ptlrpc_retain_replayable_request(struct ptlrpc_request
*req
,
2716 struct obd_import
*imp
);
2717 __u64
ptlrpc_next_xid(void);
2718 __u64
ptlrpc_sample_next_xid(void);
2719 __u64
ptlrpc_req_xid(struct ptlrpc_request
*request
);
2721 /* Set of routines to run a function in ptlrpcd context */
2722 void *ptlrpcd_alloc_work(struct obd_import
*imp
,
2723 int (*cb
)(const struct lu_env
*, void *), void *data
);
2724 void ptlrpcd_destroy_work(void *handler
);
2725 int ptlrpcd_queue_work(void *handler
);
2728 struct ptlrpc_service_buf_conf
{
2729 /* nbufs is buffers # to allocate when growing the pool */
2730 unsigned int bc_nbufs
;
2731 /* buffer size to post */
2732 unsigned int bc_buf_size
;
2733 /* portal to listed for requests on */
2734 unsigned int bc_req_portal
;
2735 /* portal of where to send replies to */
2736 unsigned int bc_rep_portal
;
2737 /* maximum request size to be accepted for this service */
2738 unsigned int bc_req_max_size
;
2739 /* maximum reply size this service can ever send */
2740 unsigned int bc_rep_max_size
;
2743 struct ptlrpc_service_thr_conf
{
2744 /* threadname should be 8 characters or less - 6 will be added on */
2746 /* threads increasing factor for each CPU */
2747 unsigned int tc_thr_factor
;
2748 /* service threads # to start on each partition while initializing */
2749 unsigned int tc_nthrs_init
;
2751 * low water of threads # upper-limit on each partition while running,
2752 * service availability may be impacted if threads number is lower
2753 * than this value. It can be ZERO if the service doesn't require
2754 * CPU affinity or there is only one partition.
2756 unsigned int tc_nthrs_base
;
2757 /* "soft" limit for total threads number */
2758 unsigned int tc_nthrs_max
;
2759 /* user specified threads number, it will be validated due to
2760 * other members of this structure. */
2761 unsigned int tc_nthrs_user
;
2762 /* set NUMA node affinity for service threads */
2763 unsigned int tc_cpu_affinity
;
2764 /* Tags for lu_context associated with service thread */
2768 struct ptlrpc_service_cpt_conf
{
2769 struct cfs_cpt_table
*cc_cptable
;
2770 /* string pattern to describe CPTs for a service */
2774 struct ptlrpc_service_conf
{
2777 /* soft watchdog timeout multiplifier to print stuck service traces */
2778 unsigned int psc_watchdog_factor
;
2779 /* buffer information */
2780 struct ptlrpc_service_buf_conf psc_buf
;
2781 /* thread information */
2782 struct ptlrpc_service_thr_conf psc_thr
;
2783 /* CPU partition information */
2784 struct ptlrpc_service_cpt_conf psc_cpt
;
2785 /* function table */
2786 struct ptlrpc_service_ops psc_ops
;
2789 /* ptlrpc/service.c */
2791 * Server-side services API. Register/unregister service, request state
2792 * management, service thread management
2796 void ptlrpc_save_lock(struct ptlrpc_request
*req
,
2797 struct lustre_handle
*lock
, int mode
, int no_ack
);
2798 void ptlrpc_commit_replies(struct obd_export
*exp
);
2799 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state
*rs
);
2800 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state
*rs
);
2801 int ptlrpc_hpreq_handler(struct ptlrpc_request
*req
);
2802 struct ptlrpc_service
*ptlrpc_register_service(
2803 struct ptlrpc_service_conf
*conf
,
2804 struct proc_dir_entry
*proc_entry
);
2805 void ptlrpc_stop_all_threads(struct ptlrpc_service
*svc
);
2807 int ptlrpc_start_threads(struct ptlrpc_service
*svc
);
2808 int ptlrpc_unregister_service(struct ptlrpc_service
*service
);
2809 int liblustre_check_services(void *arg
);
2810 void ptlrpc_daemonize(char *name
);
2811 int ptlrpc_service_health_check(struct ptlrpc_service
*);
2812 void ptlrpc_server_drop_request(struct ptlrpc_request
*req
);
2813 void ptlrpc_request_change_export(struct ptlrpc_request
*req
,
2814 struct obd_export
*export
);
2816 int ptlrpc_hr_init(void);
2817 void ptlrpc_hr_fini(void);
2821 /* ptlrpc/import.c */
2826 int ptlrpc_connect_import(struct obd_import
*imp
);
2827 int ptlrpc_init_import(struct obd_import
*imp
);
2828 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
);
2829 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
);
2830 void deuuidify(char *uuid
, const char *prefix
, char **uuid_start
,
2833 /* ptlrpc/pack_generic.c */
2834 int ptlrpc_reconnect_import(struct obd_import
*imp
);
2838 * ptlrpc msg buffer and swab interface
2842 int ptlrpc_buf_need_swab(struct ptlrpc_request
*req
, const int inout
,
2844 void ptlrpc_buf_set_swabbed(struct ptlrpc_request
*req
, const int inout
,
2846 int ptlrpc_unpack_rep_msg(struct ptlrpc_request
*req
, int len
);
2847 int ptlrpc_unpack_req_msg(struct ptlrpc_request
*req
, int len
);
2849 int lustre_msg_check_version(struct lustre_msg
*msg
, __u32 version
);
2850 void lustre_init_msg_v2(struct lustre_msg_v2
*msg
, int count
, __u32
*lens
,
2852 int lustre_pack_request(struct ptlrpc_request
*, __u32 magic
, int count
,
2853 __u32
*lens
, char **bufs
);
2854 int lustre_pack_reply(struct ptlrpc_request
*, int count
, __u32
*lens
,
2856 int lustre_pack_reply_v2(struct ptlrpc_request
*req
, int count
,
2857 __u32
*lens
, char **bufs
, int flags
);
2858 #define LPRFL_EARLY_REPLY 1
2859 int lustre_pack_reply_flags(struct ptlrpc_request
*, int count
, __u32
*lens
,
2860 char **bufs
, int flags
);
2861 int lustre_shrink_msg(struct lustre_msg
*msg
, int segment
,
2862 unsigned int newlen
, int move_data
);
2863 void lustre_free_reply_state(struct ptlrpc_reply_state
*rs
);
2864 int __lustre_unpack_msg(struct lustre_msg
*m
, int len
);
2865 int lustre_msg_hdr_size(__u32 magic
, int count
);
2866 int lustre_msg_size(__u32 magic
, int count
, __u32
*lengths
);
2867 int lustre_msg_size_v2(int count
, __u32
*lengths
);
2868 int lustre_packed_msg_size(struct lustre_msg
*msg
);
2869 int lustre_msg_early_size(void);
2870 void *lustre_msg_buf_v2(struct lustre_msg_v2
*m
, int n
, int min_size
);
2871 void *lustre_msg_buf(struct lustre_msg
*m
, int n
, int minlen
);
2872 int lustre_msg_buflen(struct lustre_msg
*m
, int n
);
2873 void lustre_msg_set_buflen(struct lustre_msg
*m
, int n
, int len
);
2874 int lustre_msg_bufcount(struct lustre_msg
*m
);
2875 char *lustre_msg_string(struct lustre_msg
*m
, int n
, int max_len
);
2876 __u32
lustre_msghdr_get_flags(struct lustre_msg
*msg
);
2877 void lustre_msghdr_set_flags(struct lustre_msg
*msg
, __u32 flags
);
2878 __u32
lustre_msg_get_flags(struct lustre_msg
*msg
);
2879 void lustre_msg_add_flags(struct lustre_msg
*msg
, int flags
);
2880 void lustre_msg_set_flags(struct lustre_msg
*msg
, int flags
);
2881 void lustre_msg_clear_flags(struct lustre_msg
*msg
, int flags
);
2882 __u32
lustre_msg_get_op_flags(struct lustre_msg
*msg
);
2883 void lustre_msg_add_op_flags(struct lustre_msg
*msg
, int flags
);
2884 void lustre_msg_set_op_flags(struct lustre_msg
*msg
, int flags
);
2885 struct lustre_handle
*lustre_msg_get_handle(struct lustre_msg
*msg
);
2886 __u32
lustre_msg_get_type(struct lustre_msg
*msg
);
2887 __u32
lustre_msg_get_version(struct lustre_msg
*msg
);
2888 void lustre_msg_add_version(struct lustre_msg
*msg
, int version
);
2889 __u32
lustre_msg_get_opc(struct lustre_msg
*msg
);
2890 __u64
lustre_msg_get_last_xid(struct lustre_msg
*msg
);
2891 __u64
lustre_msg_get_last_committed(struct lustre_msg
*msg
);
2892 __u64
*lustre_msg_get_versions(struct lustre_msg
*msg
);
2893 __u64
lustre_msg_get_transno(struct lustre_msg
*msg
);
2894 __u64
lustre_msg_get_slv(struct lustre_msg
*msg
);
2895 __u32
lustre_msg_get_limit(struct lustre_msg
*msg
);
2896 void lustre_msg_set_slv(struct lustre_msg
*msg
, __u64 slv
);
2897 void lustre_msg_set_limit(struct lustre_msg
*msg
, __u64 limit
);
2898 int lustre_msg_get_status(struct lustre_msg
*msg
);
2899 __u32
lustre_msg_get_conn_cnt(struct lustre_msg
*msg
);
2900 int lustre_msg_is_v1(struct lustre_msg
*msg
);
2901 __u32
lustre_msg_get_magic(struct lustre_msg
*msg
);
2902 __u32
lustre_msg_get_timeout(struct lustre_msg
*msg
);
2903 __u32
lustre_msg_get_service_time(struct lustre_msg
*msg
);
2904 char *lustre_msg_get_jobid(struct lustre_msg
*msg
);
2905 __u32
lustre_msg_get_cksum(struct lustre_msg
*msg
);
2906 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
2907 __u32
lustre_msg_calc_cksum(struct lustre_msg
*msg
, int compat18
);
2909 # warning "remove checksum compatibility support for b1_8"
2910 __u32
lustre_msg_calc_cksum(struct lustre_msg
*msg
);
2912 void lustre_msg_set_handle(struct lustre_msg
*msg
,struct lustre_handle
*handle
);
2913 void lustre_msg_set_type(struct lustre_msg
*msg
, __u32 type
);
2914 void lustre_msg_set_opc(struct lustre_msg
*msg
, __u32 opc
);
2915 void lustre_msg_set_last_xid(struct lustre_msg
*msg
, __u64 last_xid
);
2916 void lustre_msg_set_last_committed(struct lustre_msg
*msg
,__u64 last_committed
);
2917 void lustre_msg_set_versions(struct lustre_msg
*msg
, __u64
*versions
);
2918 void lustre_msg_set_transno(struct lustre_msg
*msg
, __u64 transno
);
2919 void lustre_msg_set_status(struct lustre_msg
*msg
, __u32 status
);
2920 void lustre_msg_set_conn_cnt(struct lustre_msg
*msg
, __u32 conn_cnt
);
2921 void ptlrpc_req_set_repsize(struct ptlrpc_request
*req
, int count
, __u32
*sizes
);
2922 void ptlrpc_request_set_replen(struct ptlrpc_request
*req
);
2923 void lustre_msg_set_timeout(struct lustre_msg
*msg
, __u32 timeout
);
2924 void lustre_msg_set_service_time(struct lustre_msg
*msg
, __u32 service_time
);
2925 void lustre_msg_set_jobid(struct lustre_msg
*msg
, char *jobid
);
2926 void lustre_msg_set_cksum(struct lustre_msg
*msg
, __u32 cksum
);
2929 lustre_shrink_reply(struct ptlrpc_request
*req
, int segment
,
2930 unsigned int newlen
, int move_data
)
2932 LASSERT(req
->rq_reply_state
);
2933 LASSERT(req
->rq_repmsg
);
2934 req
->rq_replen
= lustre_shrink_msg(req
->rq_repmsg
, segment
,
2938 #ifdef CONFIG_LUSTRE_TRANSLATE_ERRNOS
2940 static inline int ptlrpc_status_hton(int h
)
2943 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2944 * ELDLM_LOCK_ABORTED, etc.
2947 return -lustre_errno_hton(-h
);
2952 static inline int ptlrpc_status_ntoh(int n
)
2955 * See the comment in ptlrpc_status_hton().
2958 return -lustre_errno_ntoh(-n
);
2965 #define ptlrpc_status_hton(h) (h)
2966 #define ptlrpc_status_ntoh(n) (n)
2971 /** Change request phase of \a req to \a new_phase */
2973 ptlrpc_rqphase_move(struct ptlrpc_request
*req
, enum rq_phase new_phase
)
2975 if (req
->rq_phase
== new_phase
)
2978 if (new_phase
== RQ_PHASE_UNREGISTERING
) {
2979 req
->rq_next_phase
= req
->rq_phase
;
2981 atomic_inc(&req
->rq_import
->imp_unregistering
);
2984 if (req
->rq_phase
== RQ_PHASE_UNREGISTERING
) {
2986 atomic_dec(&req
->rq_import
->imp_unregistering
);
2989 DEBUG_REQ(D_INFO
, req
, "move req \"%s\" -> \"%s\"",
2990 ptlrpc_rqphase2str(req
), ptlrpc_phase2str(new_phase
));
2992 req
->rq_phase
= new_phase
;
2996 * Returns true if request \a req got early reply and hard deadline is not met
2999 ptlrpc_client_early(struct ptlrpc_request
*req
)
3001 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
3002 req
->rq_reply_deadline
> cfs_time_current_sec())
3004 return req
->rq_early
;
3008 * Returns true if we got real reply from server for this request
3011 ptlrpc_client_replied(struct ptlrpc_request
*req
)
3013 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
3014 req
->rq_reply_deadline
> cfs_time_current_sec())
3016 return req
->rq_replied
;
3019 /** Returns true if request \a req is in process of receiving server reply */
3021 ptlrpc_client_recv(struct ptlrpc_request
*req
)
3023 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
3024 req
->rq_reply_deadline
> cfs_time_current_sec())
3026 return req
->rq_receiving_reply
;
3030 ptlrpc_client_recv_or_unlink(struct ptlrpc_request
*req
)
3034 spin_lock(&req
->rq_lock
);
3035 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
3036 req
->rq_reply_deadline
> cfs_time_current_sec()) {
3037 spin_unlock(&req
->rq_lock
);
3040 rc
= req
->rq_receiving_reply
|| req
->rq_must_unlink
;
3041 spin_unlock(&req
->rq_lock
);
3046 ptlrpc_client_wake_req(struct ptlrpc_request
*req
)
3048 if (req
->rq_set
== NULL
)
3049 wake_up(&req
->rq_reply_waitq
);
3051 wake_up(&req
->rq_set
->set_waitq
);
3055 ptlrpc_rs_addref(struct ptlrpc_reply_state
*rs
)
3057 LASSERT(atomic_read(&rs
->rs_refcount
) > 0);
3058 atomic_inc(&rs
->rs_refcount
);
3062 ptlrpc_rs_decref(struct ptlrpc_reply_state
*rs
)
3064 LASSERT(atomic_read(&rs
->rs_refcount
) > 0);
3065 if (atomic_dec_and_test(&rs
->rs_refcount
))
3066 lustre_free_reply_state(rs
);
3069 /* Should only be called once per req */
3070 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request
*req
)
3072 if (req
->rq_reply_state
== NULL
)
3073 return; /* shouldn't occur */
3074 ptlrpc_rs_decref(req
->rq_reply_state
);
3075 req
->rq_reply_state
= NULL
;
3076 req
->rq_repmsg
= NULL
;
3079 static inline __u32
lustre_request_magic(struct ptlrpc_request
*req
)
3081 return lustre_msg_get_magic(req
->rq_reqmsg
);
3084 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request
*req
)
3086 switch (req
->rq_reqmsg
->lm_magic
) {
3087 case LUSTRE_MSG_MAGIC_V2
:
3088 return req
->rq_reqmsg
->lm_repsize
;
3090 LASSERTF(0, "incorrect message magic: %08x\n",
3091 req
->rq_reqmsg
->lm_magic
);
3096 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request
*req
)
3098 if (req
->rq_delay_limit
!= 0 &&
3099 cfs_time_before(cfs_time_add(req
->rq_queued_time
,
3100 cfs_time_seconds(req
->rq_delay_limit
)),
3101 cfs_time_current())) {
3107 static inline int ptlrpc_no_resend(struct ptlrpc_request
*req
)
3109 if (!req
->rq_no_resend
&& ptlrpc_send_limit_expired(req
)) {
3110 spin_lock(&req
->rq_lock
);
3111 req
->rq_no_resend
= 1;
3112 spin_unlock(&req
->rq_lock
);
3114 return req
->rq_no_resend
;
3118 ptlrpc_server_get_timeout(struct ptlrpc_service_part
*svcpt
)
3120 int at
= AT_OFF
? 0 : at_get(&svcpt
->scp_at_estimate
);
3122 return svcpt
->scp_service
->srv_watchdog_factor
*
3123 max_t(int, at
, obd_timeout
);
3126 static inline struct ptlrpc_service
*
3127 ptlrpc_req2svc(struct ptlrpc_request
*req
)
3129 LASSERT(req
->rq_rqbd
!= NULL
);
3130 return req
->rq_rqbd
->rqbd_svcpt
->scp_service
;
3133 /* ldlm/ldlm_lib.c */
3135 * Target client logic
3138 int client_obd_setup(struct obd_device
*obddev
, struct lustre_cfg
*lcfg
);
3139 int client_obd_cleanup(struct obd_device
*obddev
);
3140 int client_connect_import(const struct lu_env
*env
,
3141 struct obd_export
**exp
, struct obd_device
*obd
,
3142 struct obd_uuid
*cluuid
, struct obd_connect_data
*,
3144 int client_disconnect_export(struct obd_export
*exp
);
3145 int client_import_add_conn(struct obd_import
*imp
, struct obd_uuid
*uuid
,
3147 int client_import_del_conn(struct obd_import
*imp
, struct obd_uuid
*uuid
);
3148 int client_import_find_conn(struct obd_import
*imp
, lnet_nid_t peer
,
3149 struct obd_uuid
*uuid
);
3150 int import_set_conn_priority(struct obd_import
*imp
, struct obd_uuid
*uuid
);
3151 void client_destroy_import(struct obd_import
*imp
);
3155 /* ptlrpc/pinger.c */
3157 * Pinger API (client side only)
3160 enum timeout_event
{
3163 struct timeout_item
;
3164 typedef int (*timeout_cb_t
)(struct timeout_item
*, void *);
3165 int ptlrpc_pinger_add_import(struct obd_import
*imp
);
3166 int ptlrpc_pinger_del_import(struct obd_import
*imp
);
3167 int ptlrpc_add_timeout_client(int time
, enum timeout_event event
,
3168 timeout_cb_t cb
, void *data
,
3169 struct list_head
*obd_list
);
3170 int ptlrpc_del_timeout_client(struct list_head
*obd_list
,
3171 enum timeout_event event
);
3172 struct ptlrpc_request
* ptlrpc_prep_ping(struct obd_import
*imp
);
3173 int ptlrpc_obd_ping(struct obd_device
*obd
);
3174 void ping_evictor_start(void);
3175 void ping_evictor_stop(void);
3176 void ptlrpc_pinger_ir_up(void);
3177 void ptlrpc_pinger_ir_down(void);
3179 int ptlrpc_pinger_suppress_pings(void);
3181 /* ptlrpc daemon bind policy */
3183 /* all ptlrpcd threads are free mode */
3184 PDB_POLICY_NONE
= 1,
3185 /* all ptlrpcd threads are bound mode */
3186 PDB_POLICY_FULL
= 2,
3187 /* <free1 bound1> <free2 bound2> ... <freeN boundN> */
3188 PDB_POLICY_PAIR
= 3,
3189 /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>,
3190 * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1].
3191 * If kernel supports NUMA, pthrpcd threads are binded and
3192 * grouped by NUMA node */
3193 PDB_POLICY_NEIGHBOR
= 4,
3196 /* ptlrpc daemon load policy
3197 * It is caller's duty to specify how to push the async RPC into some ptlrpcd
3198 * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is
3199 * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd,
3200 * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner,
3201 * depends on which is scheduled firstly, to accelerate the RPC processing. */
3203 /* on the same CPU core as the caller */
3204 PDL_POLICY_SAME
= 1,
3205 /* within the same CPU partition, but not the same core as the caller */
3206 PDL_POLICY_LOCAL
= 2,
3207 /* round-robin on all CPU cores, but not the same core as the caller */
3208 PDL_POLICY_ROUND
= 3,
3209 /* the specified CPU core is preferred, but not enforced */
3210 PDL_POLICY_PREFERRED
= 4,
3213 /* ptlrpc/ptlrpcd.c */
3214 void ptlrpcd_stop(struct ptlrpcd_ctl
*pc
, int force
);
3215 void ptlrpcd_free(struct ptlrpcd_ctl
*pc
);
3216 void ptlrpcd_wake(struct ptlrpc_request
*req
);
3217 void ptlrpcd_add_req(struct ptlrpc_request
*req
, pdl_policy_t policy
, int idx
);
3218 void ptlrpcd_add_rqset(struct ptlrpc_request_set
*set
);
3219 int ptlrpcd_addref(void);
3220 void ptlrpcd_decref(void);
3222 /* ptlrpc/lproc_ptlrpc.c */
3224 * procfs output related functions
3227 const char* ll_opcode2str(__u32 opcode
);
3229 void ptlrpc_lprocfs_register_obd(struct obd_device
*obd
);
3230 void ptlrpc_lprocfs_unregister_obd(struct obd_device
*obd
);
3231 void ptlrpc_lprocfs_brw(struct ptlrpc_request
*req
, int bytes
);
3233 static inline void ptlrpc_lprocfs_register_obd(struct obd_device
*obd
) {}
3234 static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device
*obd
) {}
3235 static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request
*req
, int bytes
) {}
3239 /* ptlrpc/llog_client.c */
3240 extern struct llog_operations llog_client_ops
;