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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 /** \defgroup PtlRPC Portal RPC and networking module.
34 * PortalRPC is the layer used by rest of lustre code to achieve network
35 * communications: establish connections with corresponding export and import
36 * states, listen for a service, send and receive RPCs.
37 * PortalRPC also includes base recovery framework: packet resending and
38 * replaying, reconnections, pinger.
40 * PortalRPC utilizes LNet as its transport layer.
53 #include "../../include/linux/libcfs/libcfs.h"
54 #include "../../include/linux/lnet/nidstr.h"
55 #include "../../include/linux/lnet/api.h"
56 #include "lustre/lustre_idl.h"
57 #include "lustre_ha.h"
58 #include "lustre_sec.h"
59 #include "lustre_import.h"
60 #include "lprocfs_status.h"
61 #include "lu_object.h"
62 #include "lustre_req_layout.h"
64 #include "obd_support.h"
65 #include "lustre_ver.h"
67 /* MD flags we _always_ use */
68 #define PTLRPC_MD_OPTIONS 0
71 * Max # of bulk operations in one request.
72 * In order for the client and server to properly negotiate the maximum
73 * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two
74 * value. The client is free to limit the actual RPC size for any bulk
75 * transfer via cl_max_pages_per_rpc to some non-power-of-two value.
77 #define PTLRPC_BULK_OPS_BITS 2
78 #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS)
80 * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and
81 * should not be used on the server at all. Otherwise, it imposes a
82 * protocol limitation on the maximum RPC size that can be used by any
83 * RPC sent to that server in the future. Instead, the server should
84 * use the negotiated per-client ocd_brw_size to determine the bulk
87 #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1))
90 * Define maxima for bulk I/O.
92 * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT
93 * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the
94 * currently supported maximum between peers at connect via ocd_brw_size.
96 #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS)
97 #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS)
98 #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_SHIFT)
100 #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS)
101 #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
102 #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_SHIFT)
103 #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE
104 #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_SHIFT)
105 #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS)
107 /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */
108 # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0)
109 # error "PTLRPC_MAX_BRW_PAGES isn't a power of two"
111 # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_SIZE))
112 # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_SIZE"
114 # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT)
115 # error "PTLRPC_MAX_BRW_SIZE too big"
117 # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT)
118 # error "PTLRPC_MAX_BRW_PAGES too big"
121 #define PTLRPC_NTHRS_INIT 2
126 * Constants determine how memory is used to buffer incoming service requests.
128 * ?_NBUFS # buffers to allocate when growing the pool
129 * ?_BUFSIZE # bytes in a single request buffer
130 * ?_MAXREQSIZE # maximum request service will receive
132 * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk
133 * of ?_NBUFS is added to the pool.
135 * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are
136 * considered full when less than ?_MAXREQSIZE is left in them.
141 * Constants determine how threads are created for ptlrpc service.
143 * ?_NTHRS_INIT # threads to create for each service partition on
144 * initializing. If it's non-affinity service and
145 * there is only one partition, it's the overall #
146 * threads for the service while initializing.
147 * ?_NTHRS_BASE # threads should be created at least for each
148 * ptlrpc partition to keep the service healthy.
149 * It's the low-water mark of threads upper-limit
150 * for each partition.
151 * ?_THR_FACTOR # threads can be added on threads upper-limit for
152 * each CPU core. This factor is only for reference,
153 * we might decrease value of factor if number of cores
154 * per CPT is above a limit.
155 * ?_NTHRS_MAX # overall threads can be created for a service,
156 * it's a soft limit because if service is running
157 * on machine with hundreds of cores and tens of
158 * CPU partitions, we need to guarantee each partition
159 * has ?_NTHRS_BASE threads, which means total threads
160 * will be ?_NTHRS_BASE * number_of_cpts which can
161 * exceed ?_NTHRS_MAX.
165 * #define MDS_NTHRS_INIT 2
166 * #define MDS_NTHRS_BASE 64
167 * #define MDS_NTHRS_FACTOR 8
168 * #define MDS_NTHRS_MAX 1024
171 * ---------------------------------------------------------------------
172 * Server(A) has 16 cores, user configured it to 4 partitions so each
173 * partition has 4 cores, then actual number of service threads on each
175 * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96
177 * Total number of threads for the service is:
178 * 96 * partitions(4) = 384
181 * ---------------------------------------------------------------------
182 * Server(B) has 32 cores, user configured it to 4 partitions so each
183 * partition has 8 cores, then actual number of service threads on each
185 * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128
187 * Total number of threads for the service is:
188 * 128 * partitions(4) = 512
191 * ---------------------------------------------------------------------
192 * Server(B) has 96 cores, user configured it to 8 partitions so each
193 * partition has 12 cores, then actual number of service threads on each
195 * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160
197 * Total number of threads for the service is:
198 * 160 * partitions(8) = 1280
200 * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number
201 * as upper limit of threads number for each partition:
202 * MDS_NTHRS_MAX(1024) / partitions(8) = 128
205 * ---------------------------------------------------------------------
206 * Server(C) have a thousand of cores and user configured it to 32 partitions
207 * MDS_NTHRS_BASE(64) * 32 = 2048
209 * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need
210 * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads
211 * to keep service healthy, so total number of threads will just be 2048.
213 * NB: we don't suggest to choose server with that many cores because backend
214 * filesystem itself, buffer cache, or underlying network stack might
215 * have some SMP scalability issues at that large scale.
217 * If user already has a fat machine with hundreds or thousands of cores,
218 * there are two choices for configuration:
219 * a) create CPU table from subset of all CPUs and run Lustre on
221 * b) bind service threads on a few partitions, see modparameters of
222 * MDS and OSS for details
224 * NB: these calculations (and examples below) are simplified to help
225 * understanding, the real implementation is a little more complex,
226 * please see ptlrpc_server_nthreads_check() for details.
231 * LDLM threads constants:
233 * Given 8 as factor and 24 as base threads number
236 * On 4-core machine we will have 24 + 8 * 4 = 56 threads.
239 * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56
240 * threads for each partition and total threads number will be 112.
243 * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
244 * threads for each partition to keep service healthy, so total threads
245 * number should be 24 * 8 = 192.
247 * So with these constants, threads number will be at the similar level
248 * of old versions, unless target machine has over a hundred cores
250 #define LDLM_THR_FACTOR 8
251 #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT
252 #define LDLM_NTHRS_BASE 24
253 #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128)
255 #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT
256 #define LDLM_CLIENT_NBUFS 1
257 #define LDLM_SERVER_NBUFS 64
258 #define LDLM_BUFSIZE (8 * 1024)
259 #define LDLM_MAXREQSIZE (5 * 1024)
260 #define LDLM_MAXREPSIZE (1024)
262 #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */
264 #define OST_MAXREQSIZE (5 * 1024)
266 /* Macro to hide a typecast. */
267 #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args)
270 * Structure to single define portal connection.
272 struct ptlrpc_connection
{
273 /** linkage for connections hash table */
274 struct hlist_node c_hash
;
275 /** Our own lnet nid for this connection */
277 /** Remote side nid for this connection */
278 lnet_process_id_t c_peer
;
279 /** UUID of the other side */
280 struct obd_uuid c_remote_uuid
;
281 /** reference counter for this connection */
285 /** Client definition for PortalRPC */
286 struct ptlrpc_client
{
287 /** What lnet portal does this client send messages to by default */
288 __u32 cli_request_portal
;
289 /** What portal do we expect replies on */
290 __u32 cli_reply_portal
;
291 /** Name of the client */
295 /** state flags of requests */
296 /* XXX only ones left are those used by the bulk descs as well! */
297 #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */
298 #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */
300 #define REQ_MAX_ACK_LOCKS 8
302 union ptlrpc_async_args
{
304 * Scratchpad for passing args to completion interpreter. Users
305 * cast to the struct of their choosing, and CLASSERT that this is
306 * big enough. For _tons_ of context, kmalloc a struct and store
307 * a pointer to it here. The pointer_arg ensures this struct is at
308 * least big enough for that.
310 void *pointer_arg
[11];
314 struct ptlrpc_request_set
;
315 typedef int (*set_interpreter_func
)(struct ptlrpc_request_set
*, void *, int);
316 typedef int (*set_producer_func
)(struct ptlrpc_request_set
*, void *);
319 * Definition of request set structure.
320 * Request set is a list of requests (not necessary to the same target) that
321 * once populated with RPCs could be sent in parallel.
322 * There are two kinds of request sets. General purpose and with dedicated
323 * serving thread. Example of the latter is ptlrpcd set.
324 * For general purpose sets once request set started sending it is impossible
325 * to add new requests to such set.
326 * Provides a way to call "completion callbacks" when all requests in the set
329 struct ptlrpc_request_set
{
330 atomic_t set_refcount
;
331 /** number of in queue requests */
332 atomic_t set_new_count
;
333 /** number of uncompleted requests */
334 atomic_t set_remaining
;
335 /** wait queue to wait on for request events */
336 wait_queue_head_t set_waitq
;
337 wait_queue_head_t
*set_wakeup_ptr
;
338 /** List of requests in the set */
339 struct list_head set_requests
;
341 * List of completion callbacks to be called when the set is completed
342 * This is only used if \a set_interpret is NULL.
343 * Links struct ptlrpc_set_cbdata.
345 struct list_head set_cblist
;
346 /** Completion callback, if only one. */
347 set_interpreter_func set_interpret
;
348 /** opaq argument passed to completion \a set_interpret callback. */
351 * Lock for \a set_new_requests manipulations
352 * locked so that any old caller can communicate requests to
353 * the set holder who can then fold them into the lock-free set
355 spinlock_t set_new_req_lock
;
356 /** List of new yet unsent requests. Only used with ptlrpcd now. */
357 struct list_head set_new_requests
;
359 /** rq_status of requests that have been freed already */
361 /** Additional fields used by the flow control extension */
362 /** Maximum number of RPCs in flight */
363 int set_max_inflight
;
364 /** Callback function used to generate RPCs */
365 set_producer_func set_producer
;
366 /** opaq argument passed to the producer callback */
367 void *set_producer_arg
;
371 * Description of a single ptrlrpc_set callback
373 struct ptlrpc_set_cbdata
{
374 /** List linkage item */
375 struct list_head psc_item
;
376 /** Pointer to interpreting function */
377 set_interpreter_func psc_interpret
;
378 /** Opaq argument to pass to the callback */
382 struct ptlrpc_bulk_desc
;
383 struct ptlrpc_service_part
;
384 struct ptlrpc_service
;
387 * ptlrpc callback & work item stuff
389 struct ptlrpc_cb_id
{
390 void (*cbid_fn
)(lnet_event_t
*ev
); /* specific callback fn */
391 void *cbid_arg
; /* additional arg */
394 /** Maximum number of locks to fit into reply state */
395 #define RS_MAX_LOCKS 8
399 * Structure to define reply state on the server
400 * Reply state holds various reply message information. Also for "difficult"
401 * replies (rep-ack case) we store the state after sending reply and wait
402 * for the client to acknowledge the reception. In these cases locks could be
403 * added to the state for replay/failover consistency guarantees.
405 struct ptlrpc_reply_state
{
406 /** Callback description */
407 struct ptlrpc_cb_id rs_cb_id
;
408 /** Linkage for list of all reply states in a system */
409 struct list_head rs_list
;
410 /** Linkage for list of all reply states on same export */
411 struct list_head rs_exp_list
;
412 /** Linkage for list of all reply states for same obd */
413 struct list_head rs_obd_list
;
415 struct list_head rs_debug_list
;
417 /** A spinlock to protect the reply state flags */
419 /** Reply state flags */
420 unsigned long rs_difficult
:1; /* ACK/commit stuff */
421 unsigned long rs_no_ack
:1; /* no ACK, even for
424 unsigned long rs_scheduled
:1; /* being handled? */
425 unsigned long rs_scheduled_ever
:1;/* any schedule attempts? */
426 unsigned long rs_handled
:1; /* been handled yet? */
427 unsigned long rs_on_net
:1; /* reply_out_callback pending? */
428 unsigned long rs_prealloc
:1; /* rs from prealloc list */
429 unsigned long rs_committed
:1;/* the transaction was committed
430 * and the rs was dispatched
432 /** Size of the state */
436 /** Transaction number */
440 struct obd_export
*rs_export
;
441 struct ptlrpc_service_part
*rs_svcpt
;
442 /** Lnet metadata handle for the reply */
443 lnet_handle_md_t rs_md_h
;
444 atomic_t rs_refcount
;
446 /** Context for the service thread */
447 struct ptlrpc_svc_ctx
*rs_svc_ctx
;
448 /** Reply buffer (actually sent to the client), encoded if needed */
449 struct lustre_msg
*rs_repbuf
; /* wrapper */
450 /** Size of the reply buffer */
451 int rs_repbuf_len
; /* wrapper buf length */
452 /** Size of the reply message */
453 int rs_repdata_len
; /* wrapper msg length */
455 * Actual reply message. Its content is encrypted (if needed) to
456 * produce reply buffer for actual sending. In simple case
457 * of no network encryption we just set \a rs_repbuf to \a rs_msg
459 struct lustre_msg
*rs_msg
; /* reply message */
461 /** Number of locks awaiting client ACK */
463 /** Handles of locks awaiting client reply ACK */
464 struct lustre_handle rs_locks
[RS_MAX_LOCKS
];
465 /** Lock modes of locks in \a rs_locks */
466 enum ldlm_mode rs_modes
[RS_MAX_LOCKS
];
469 struct ptlrpc_thread
;
473 RQ_PHASE_NEW
= 0xebc0de00,
474 RQ_PHASE_RPC
= 0xebc0de01,
475 RQ_PHASE_BULK
= 0xebc0de02,
476 RQ_PHASE_INTERPRET
= 0xebc0de03,
477 RQ_PHASE_COMPLETE
= 0xebc0de04,
478 RQ_PHASE_UNREGISTERING
= 0xebc0de05,
479 RQ_PHASE_UNDEFINED
= 0xebc0de06
482 /** Type of request interpreter call-back */
483 typedef int (*ptlrpc_interpterer_t
)(const struct lu_env
*env
,
484 struct ptlrpc_request
*req
,
488 * Definition of request pool structure.
489 * The pool is used to store empty preallocated requests for the case
490 * when we would actually need to send something without performing
491 * any allocations (to avoid e.g. OOM).
493 struct ptlrpc_request_pool
{
494 /** Locks the list */
496 /** list of ptlrpc_request structs */
497 struct list_head prp_req_list
;
498 /** Maximum message size that would fit into a request from this pool */
500 /** Function to allocate more requests for this pool */
501 int (*prp_populate
)(struct ptlrpc_request_pool
*, int);
510 * \defgroup nrs Network Request Scheduler
513 struct ptlrpc_nrs_policy
;
514 struct ptlrpc_nrs_resource
;
515 struct ptlrpc_nrs_request
;
518 * NRS control operations.
520 * These are common for all policies.
522 enum ptlrpc_nrs_ctl
{
524 * Not a valid opcode.
526 PTLRPC_NRS_CTL_INVALID
,
528 * Activate the policy.
530 PTLRPC_NRS_CTL_START
,
532 * Reserved for multiple primary policies, which may be a possibility
537 * Policies can start using opcodes from this value and onwards for
538 * their own purposes; the assigned value itself is arbitrary.
540 PTLRPC_NRS_CTL_1ST_POL_SPEC
= 0x20,
544 * ORR policy operations
547 NRS_CTL_ORR_RD_QUANTUM
= PTLRPC_NRS_CTL_1ST_POL_SPEC
,
548 NRS_CTL_ORR_WR_QUANTUM
,
549 NRS_CTL_ORR_RD_OFF_TYPE
,
550 NRS_CTL_ORR_WR_OFF_TYPE
,
551 NRS_CTL_ORR_RD_SUPP_REQ
,
552 NRS_CTL_ORR_WR_SUPP_REQ
,
556 * NRS policy operations.
558 * These determine the behaviour of a policy, and are called in response to
561 struct ptlrpc_nrs_pol_ops
{
563 * Called during policy registration; this operation is optional.
565 * \param[in,out] policy The policy being initialized
567 int (*op_policy_init
) (struct ptlrpc_nrs_policy
*policy
);
569 * Called during policy unregistration; this operation is optional.
571 * \param[in,out] policy The policy being unregistered/finalized
573 void (*op_policy_fini
) (struct ptlrpc_nrs_policy
*policy
);
575 * Called when activating a policy via lprocfs; policies allocate and
576 * initialize their resources here; this operation is optional.
578 * \param[in,out] policy The policy being started
580 * \see nrs_policy_start_locked()
582 int (*op_policy_start
) (struct ptlrpc_nrs_policy
*policy
);
584 * Called when deactivating a policy via lprocfs; policies deallocate
585 * their resources here; this operation is optional
587 * \param[in,out] policy The policy being stopped
589 * \see nrs_policy_stop0()
591 void (*op_policy_stop
) (struct ptlrpc_nrs_policy
*policy
);
593 * Used for policy-specific operations; i.e. not generic ones like
594 * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous
595 * to an ioctl; this operation is optional.
597 * \param[in,out] policy The policy carrying out operation \a opc
598 * \param[in] opc The command operation being carried out
599 * \param[in,out] arg An generic buffer for communication between the
600 * user and the control operation
605 * \see ptlrpc_nrs_policy_control()
607 int (*op_policy_ctl
) (struct ptlrpc_nrs_policy
*policy
,
608 enum ptlrpc_nrs_ctl opc
, void *arg
);
611 * Called when obtaining references to the resources of the resource
612 * hierarchy for a request that has arrived for handling at the PTLRPC
613 * service. Policies should return -ve for requests they do not wish
614 * to handle. This operation is mandatory.
616 * \param[in,out] policy The policy we're getting resources for.
617 * \param[in,out] nrq The request we are getting resources for.
618 * \param[in] parent The parent resource of the resource being
619 * requested; set to NULL if none.
620 * \param[out] resp The resource is to be returned here; the
621 * fallback policy in an NRS head should
622 * \e always return a non-NULL pointer value.
623 * \param[in] moving_req When set, signifies that this is an attempt
624 * to obtain resources for a request being moved
625 * to the high-priority NRS head by
626 * ldlm_lock_reorder_req().
627 * This implies two things:
628 * 1. We are under obd_export::exp_rpc_lock and
629 * so should not sleep.
630 * 2. We should not perform non-idempotent or can
631 * skip performing idempotent operations that
632 * were carried out when resources were first
633 * taken for the request when it was initialized
634 * in ptlrpc_nrs_req_initialize().
636 * \retval 0, +ve The level of the returned resource in the resource
637 * hierarchy; currently only 0 (for a non-leaf resource)
638 * and 1 (for a leaf resource) are supported by the
642 * \see ptlrpc_nrs_req_initialize()
643 * \see ptlrpc_nrs_hpreq_add_nolock()
645 int (*op_res_get
) (struct ptlrpc_nrs_policy
*policy
,
646 struct ptlrpc_nrs_request
*nrq
,
647 const struct ptlrpc_nrs_resource
*parent
,
648 struct ptlrpc_nrs_resource
**resp
,
651 * Called when releasing references taken for resources in the resource
652 * hierarchy for the request; this operation is optional.
654 * \param[in,out] policy The policy the resource belongs to
655 * \param[in] res The resource to be freed
657 * \see ptlrpc_nrs_req_finalize()
658 * \see ptlrpc_nrs_hpreq_add_nolock()
660 void (*op_res_put
) (struct ptlrpc_nrs_policy
*policy
,
661 const struct ptlrpc_nrs_resource
*res
);
664 * Obtains a request for handling from the policy, and optionally
665 * removes the request from the policy; this operation is mandatory.
667 * \param[in,out] policy The policy to poll
668 * \param[in] peek When set, signifies that we just want to
669 * examine the request, and not handle it, so the
670 * request is not removed from the policy.
671 * \param[in] force When set, it will force a policy to return a
672 * request if it has one queued.
674 * \retval NULL No request available for handling
675 * \retval valid-pointer The request polled for handling
677 * \see ptlrpc_nrs_req_get_nolock()
679 struct ptlrpc_nrs_request
*
680 (*op_req_get
) (struct ptlrpc_nrs_policy
*policy
, bool peek
,
683 * Called when attempting to add a request to a policy for later
684 * handling; this operation is mandatory.
686 * \param[in,out] policy The policy on which to enqueue \a nrq
687 * \param[in,out] nrq The request to enqueue
692 * \see ptlrpc_nrs_req_add_nolock()
694 int (*op_req_enqueue
) (struct ptlrpc_nrs_policy
*policy
,
695 struct ptlrpc_nrs_request
*nrq
);
697 * Removes a request from the policy's set of pending requests. Normally
698 * called after a request has been polled successfully from the policy
699 * for handling; this operation is mandatory.
701 * \param[in,out] policy The policy the request \a nrq belongs to
702 * \param[in,out] nrq The request to dequeue
704 void (*op_req_dequeue
) (struct ptlrpc_nrs_policy
*policy
,
705 struct ptlrpc_nrs_request
*nrq
);
707 * Called after the request being carried out. Could be used for
708 * job/resource control; this operation is optional.
710 * \param[in,out] policy The policy which is stopping to handle request
712 * \param[in,out] nrq The request
714 * \pre assert_spin_locked(&svcpt->scp_req_lock)
716 * \see ptlrpc_nrs_req_stop_nolock()
718 void (*op_req_stop
) (struct ptlrpc_nrs_policy
*policy
,
719 struct ptlrpc_nrs_request
*nrq
);
721 * Registers the policy's lprocfs interface with a PTLRPC service.
723 * \param[in] svc The service
728 int (*op_lprocfs_init
) (struct ptlrpc_service
*svc
);
730 * Unegisters the policy's lprocfs interface with a PTLRPC service.
732 * In cases of failed policy registration in
733 * \e ptlrpc_nrs_policy_register(), this function may be called for a
734 * service which has not registered the policy successfully, so
735 * implementations of this method should make sure their operations are
736 * safe in such cases.
738 * \param[in] svc The service
740 void (*op_lprocfs_fini
) (struct ptlrpc_service
*svc
);
746 enum nrs_policy_flags
{
748 * Fallback policy, use this flag only on a single supported policy per
749 * service. The flag cannot be used on policies that use
750 * \e PTLRPC_NRS_FL_REG_EXTERN
752 PTLRPC_NRS_FL_FALLBACK
= (1 << 0),
754 * Start policy immediately after registering.
756 PTLRPC_NRS_FL_REG_START
= (1 << 1),
758 * This is a policy registering from a module different to the one NRS
759 * core ships in (currently ptlrpc).
761 PTLRPC_NRS_FL_REG_EXTERN
= (1 << 2),
767 * Denotes whether an NRS instance is for handling normal or high-priority
768 * RPCs, or whether an operation pertains to one or both of the NRS instances
771 enum ptlrpc_nrs_queue_type
{
772 PTLRPC_NRS_QUEUE_REG
= (1 << 0),
773 PTLRPC_NRS_QUEUE_HP
= (1 << 1),
774 PTLRPC_NRS_QUEUE_BOTH
= (PTLRPC_NRS_QUEUE_REG
| PTLRPC_NRS_QUEUE_HP
)
780 * A PTLRPC service has at least one NRS head instance for handling normal
781 * priority RPCs, and may optionally have a second NRS head instance for
782 * handling high-priority RPCs. Each NRS head maintains a list of available
783 * policies, of which one and only one policy is acting as the fallback policy,
784 * and optionally a different policy may be acting as the primary policy. For
785 * all RPCs handled by this NRS head instance, NRS core will first attempt to
786 * enqueue the RPC using the primary policy (if any). The fallback policy is
787 * used in the following cases:
788 * - when there was no primary policy in the
789 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request
791 * - when the primary policy that was at the
792 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
793 * RPC was initialized, denoted it did not wish, or for some other reason was
794 * not able to handle the request, by returning a non-valid NRS resource
796 * - when the primary policy that was at the
797 * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the
798 * RPC was initialized, fails later during the request enqueueing stage.
800 * \see nrs_resource_get_safe()
801 * \see nrs_request_enqueue()
805 /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */
807 * List of registered policies
809 struct list_head nrs_policy_list
;
811 * List of policies with queued requests. Policies that have any
812 * outstanding requests are queued here, and this list is queried
813 * in a round-robin manner from NRS core when obtaining a request
814 * for handling. This ensures that requests from policies that at some
815 * point transition away from the
816 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained.
818 struct list_head nrs_policy_queued
;
820 * Service partition for this NRS head
822 struct ptlrpc_service_part
*nrs_svcpt
;
824 * Primary policy, which is the preferred policy for handling RPCs
826 struct ptlrpc_nrs_policy
*nrs_policy_primary
;
828 * Fallback policy, which is the backup policy for handling RPCs
830 struct ptlrpc_nrs_policy
*nrs_policy_fallback
;
832 * This NRS head handles either HP or regular requests
834 enum ptlrpc_nrs_queue_type nrs_queue_type
;
836 * # queued requests from all policies in this NRS head
838 unsigned long nrs_req_queued
;
840 * # scheduled requests from all policies in this NRS head
842 unsigned long nrs_req_started
;
844 * # policies on this NRS
846 unsigned nrs_num_pols
;
848 * This NRS head is in progress of starting a policy
850 unsigned nrs_policy_starting
:1;
852 * In progress of shutting down the whole NRS head; used during
855 unsigned nrs_stopping
:1;
858 #define NRS_POL_NAME_MAX 16
860 struct ptlrpc_nrs_pol_desc
;
863 * Service compatibility predicate; this determines whether a policy is adequate
864 * for handling RPCs of a particular PTLRPC service.
866 * XXX:This should give the same result during policy registration and
867 * unregistration, and for all partitions of a service; so the result should not
868 * depend on temporal service or other properties, that may influence the
871 typedef bool (*nrs_pol_desc_compat_t
) (const struct ptlrpc_service
*svc
,
872 const struct ptlrpc_nrs_pol_desc
*desc
);
874 struct ptlrpc_nrs_pol_conf
{
876 * Human-readable policy name
878 char nc_name
[NRS_POL_NAME_MAX
];
880 * NRS operations for this policy
882 const struct ptlrpc_nrs_pol_ops
*nc_ops
;
884 * Service compatibility predicate
886 nrs_pol_desc_compat_t nc_compat
;
888 * Set for policies that support a single ptlrpc service, i.e. ones that
889 * have \a pd_compat set to nrs_policy_compat_one(). The variable value
890 * depicts the name of the single service that such policies are
893 const char *nc_compat_svc_name
;
895 * Owner module for this policy descriptor; policies registering from a
896 * different module to the one the NRS framework is held within
897 * (currently ptlrpc), should set this field to THIS_MODULE.
899 struct module
*nc_owner
;
901 * Policy registration flags; a bitmask of \e nrs_policy_flags
907 * NRS policy registering descriptor
909 * Is used to hold a description of a policy that can be passed to NRS core in
910 * order to register the policy with NRS heads in different PTLRPC services.
912 struct ptlrpc_nrs_pol_desc
{
914 * Human-readable policy name
916 char pd_name
[NRS_POL_NAME_MAX
];
918 * Link into nrs_core::nrs_policies
920 struct list_head pd_list
;
922 * NRS operations for this policy
924 const struct ptlrpc_nrs_pol_ops
*pd_ops
;
926 * Service compatibility predicate
928 nrs_pol_desc_compat_t pd_compat
;
930 * Set for policies that are compatible with only one PTLRPC service.
932 * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name
934 const char *pd_compat_svc_name
;
936 * Owner module for this policy descriptor.
938 * We need to hold a reference to the module whenever we might make use
939 * of any of the module's contents, i.e.
940 * - If one or more instances of the policy are at a state where they
941 * might be handling a request, i.e.
942 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
943 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to
944 * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference
945 * is taken on the module when
946 * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it
947 * becomes 0, so that we hold only one reference to the module maximum
950 * We do not need to hold a reference to the module, even though we
951 * might use code and data from the module, in the following cases:
952 * - During external policy registration, because this should happen in
953 * the module's init() function, in which case the module is safe from
954 * removal because a reference is being held on the module by the
955 * kernel, and iirc kmod (and I guess module-init-tools also) will
956 * serialize any racing processes properly anyway.
957 * - During external policy unregistration, because this should happen
958 * in a module's exit() function, and any attempts to start a policy
959 * instance would need to take a reference on the module, and this is
960 * not possible once we have reached the point where the exit()
962 * - During service registration and unregistration, as service setup
963 * and cleanup, and policy registration, unregistration and policy
964 * instance starting, are serialized by \e nrs_core::nrs_mutex, so
965 * as long as users adhere to the convention of registering policies
966 * in init() and unregistering them in module exit() functions, there
967 * should not be a race between these operations.
968 * - During any policy-specific lprocfs operations, because a reference
969 * is held by the kernel on a proc entry that has been entered by a
970 * syscall, so as long as proc entries are removed during unregistration time,
971 * then unregistration and lprocfs operations will be properly
974 struct module
*pd_owner
;
976 * Bitmask of \e nrs_policy_flags
980 * # of references on this descriptor
988 * Policies transition from one state to the other during their lifetime
990 enum ptlrpc_nrs_pol_state
{
992 * Not a valid policy state.
994 NRS_POL_STATE_INVALID
,
996 * Policies are at this state either at the start of their life, or
997 * transition here when the user selects a different policy to act
998 * as the primary one.
1000 NRS_POL_STATE_STOPPED
,
1002 * Policy is progress of stopping
1004 NRS_POL_STATE_STOPPING
,
1006 * Policy is in progress of starting
1008 NRS_POL_STATE_STARTING
,
1010 * A policy is in this state in two cases:
1011 * - it is the fallback policy, which is always in this state.
1012 * - it has been activated by the user; i.e. it is the primary policy,
1014 NRS_POL_STATE_STARTED
,
1018 * NRS policy information
1020 * Used for obtaining information for the status of a policy via lprocfs
1022 struct ptlrpc_nrs_pol_info
{
1026 char pi_name
[NRS_POL_NAME_MAX
];
1028 * Current policy state
1030 enum ptlrpc_nrs_pol_state pi_state
;
1032 * # RPCs enqueued for later dispatching by the policy
1036 * # RPCs started for dispatch by the policy
1038 long pi_req_started
;
1040 * Is this a fallback policy?
1042 unsigned pi_fallback
:1;
1048 * There is one instance of this for each policy in each NRS head of each
1049 * PTLRPC service partition.
1051 struct ptlrpc_nrs_policy
{
1053 * Linkage into the NRS head's list of policies,
1054 * ptlrpc_nrs:nrs_policy_list
1056 struct list_head pol_list
;
1058 * Linkage into the NRS head's list of policies with enqueued
1059 * requests ptlrpc_nrs:nrs_policy_queued
1061 struct list_head pol_list_queued
;
1063 * Current state of this policy
1065 enum ptlrpc_nrs_pol_state pol_state
;
1067 * Bitmask of nrs_policy_flags
1071 * # RPCs enqueued for later dispatching by the policy
1073 long pol_req_queued
;
1075 * # RPCs started for dispatch by the policy
1077 long pol_req_started
;
1079 * Usage Reference count taken on the policy instance
1083 * The NRS head this policy has been created at
1085 struct ptlrpc_nrs
*pol_nrs
;
1087 * Private policy data; varies by policy type
1091 * Policy descriptor for this policy instance.
1093 struct ptlrpc_nrs_pol_desc
*pol_desc
;
1099 * Resources are embedded into two types of NRS entities:
1100 * - Inside NRS policies, in the policy's private data in
1101 * ptlrpc_nrs_policy::pol_private
1102 * - In objects that act as prime-level scheduling entities in different NRS
1103 * policies; e.g. on a policy that performs round robin or similar order
1104 * scheduling across client NIDs, there would be one NRS resource per unique
1105 * client NID. On a policy which performs round robin scheduling across
1106 * backend filesystem objects, there would be one resource associated with
1107 * each of the backend filesystem objects partaking in the scheduling
1108 * performed by the policy.
1110 * NRS resources share a parent-child relationship, in which resources embedded
1111 * in policy instances are the parent entities, with all scheduling entities
1112 * a policy schedules across being the children, thus forming a simple resource
1113 * hierarchy. This hierarchy may be extended with one or more levels in the
1114 * future if the ability to have more than one primary policy is added.
1116 * Upon request initialization, references to the then active NRS policies are
1117 * taken and used to later handle the dispatching of the request with one of
1120 * \see nrs_resource_get_safe()
1121 * \see ptlrpc_nrs_req_add()
1123 struct ptlrpc_nrs_resource
{
1125 * This NRS resource's parent; is NULL for resources embedded in NRS
1126 * policy instances; i.e. those are top-level ones.
1128 struct ptlrpc_nrs_resource
*res_parent
;
1130 * The policy associated with this resource.
1132 struct ptlrpc_nrs_policy
*res_policy
;
1145 * This policy is a logical wrapper around previous, non-NRS functionality.
1146 * It dispatches RPCs in the same order as they arrive from the network. This
1147 * policy is currently used as the fallback policy, and the only enabled policy
1148 * on all NRS heads of all PTLRPC service partitions.
1153 * Private data structure for the FIFO policy
1155 struct nrs_fifo_head
{
1157 * Resource object for policy instance.
1159 struct ptlrpc_nrs_resource fh_res
;
1161 * List of queued requests.
1163 struct list_head fh_list
;
1165 * For debugging purposes.
1170 struct nrs_fifo_req
{
1171 struct list_head fr_list
;
1180 * Instances of this object exist embedded within ptlrpc_request; the main
1181 * purpose of this object is to hold references to the request's resources
1182 * for the lifetime of the request, and to hold properties that policies use
1183 * use for determining the request's scheduling priority.
1185 struct ptlrpc_nrs_request
{
1187 * The request's resource hierarchy.
1189 struct ptlrpc_nrs_resource
*nr_res_ptrs
[NRS_RES_MAX
];
1191 * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the
1192 * policy that was used to enqueue the request.
1194 * \see nrs_request_enqueue()
1196 unsigned nr_res_idx
;
1197 unsigned nr_initialized
:1;
1198 unsigned nr_enqueued
:1;
1199 unsigned nr_started
:1;
1200 unsigned nr_finalized
:1;
1203 * Policy-specific fields, used for determining a request's scheduling
1204 * priority, and other supporting functionality.
1208 * Fields for the FIFO policy
1210 struct nrs_fifo_req fifo
;
1213 * Externally-registering policies may want to use this to allocate
1214 * their own request properties.
1222 * Basic request prioritization operations structure.
1223 * The whole idea is centered around locks and RPCs that might affect locks.
1224 * When a lock is contended we try to give priority to RPCs that might lead
1225 * to fastest release of that lock.
1226 * Currently only implemented for OSTs only in a way that makes all
1227 * IO and truncate RPCs that are coming from a locked region where a lock is
1228 * contended a priority over other requests.
1230 struct ptlrpc_hpreq_ops
{
1232 * Check if the lock handle of the given lock is the same as
1233 * taken from the request.
1235 int (*hpreq_lock_match
)(struct ptlrpc_request
*, struct ldlm_lock
*);
1237 * Check if the request is a high priority one.
1239 int (*hpreq_check
)(struct ptlrpc_request
*);
1241 * Called after the request has been handled.
1243 void (*hpreq_fini
)(struct ptlrpc_request
*);
1247 * Represents remote procedure call.
1249 * This is a staple structure used by everybody wanting to send a request
1252 struct ptlrpc_request
{
1253 /* Request type: one of PTL_RPC_MSG_* */
1255 /** Result of request processing */
1258 * Linkage item through which this request is included into
1259 * sending/delayed lists on client and into rqbd list on server
1261 struct list_head rq_list
;
1263 * Server side list of incoming unserved requests sorted by arrival
1264 * time. Traversed from time to time to notice about to expire
1265 * requests and sent back "early replies" to clients to let them
1266 * know server is alive and well, just very busy to service their
1269 struct list_head rq_timed_list
;
1270 /** server-side history, used for debugging purposes. */
1271 struct list_head rq_history_list
;
1272 /** server-side per-export list */
1273 struct list_head rq_exp_list
;
1274 /** server-side hp handlers */
1275 struct ptlrpc_hpreq_ops
*rq_ops
;
1277 /** initial thread servicing this request */
1278 struct ptlrpc_thread
*rq_svc_thread
;
1280 /** history sequence # */
1281 __u64 rq_history_seq
;
1285 /** stub for NRS request */
1286 struct ptlrpc_nrs_request rq_nrq
;
1288 /** the index of service's srv_at_array into which request is linked */
1290 /** Lock to protect request flags and some other important bits, like
1294 /** client-side flags are serialized by rq_lock */
1295 unsigned int rq_intr
:1, rq_replied
:1, rq_err
:1,
1296 rq_timedout
:1, rq_resend
:1, rq_restart
:1,
1298 * when ->rq_replay is set, request is kept by the client even
1299 * after server commits corresponding transaction. This is
1300 * used for operations that require sequence of multiple
1301 * requests to be replayed. The only example currently is file
1302 * open/close. When last request in such a sequence is
1303 * committed, ->rq_replay is cleared on all requests in the
1307 rq_no_resend
:1, rq_waiting
:1, rq_receiving_reply
:1,
1308 rq_no_delay
:1, rq_net_err
:1, rq_wait_ctx
:1,
1310 rq_req_unlink
:1, rq_reply_unlink
:1,
1311 rq_memalloc
:1, /* req originated from "kswapd" */
1312 /* server-side flags */
1313 rq_packed_final
:1, /* packed final reply */
1314 rq_hp
:1, /* high priority RPC */
1315 rq_at_linked
:1, /* link into service's srv_at_array */
1316 rq_reply_truncate
:1,
1318 /* whether the "rq_set" is a valid one */
1320 rq_generation_set
:1,
1321 /* do not resend request on -EINPROGRESS */
1322 rq_no_retry_einprogress
:1,
1323 /* allow the req to be sent if the import is in recovery
1327 /* bulk request, sent to server, but uncommitted */
1330 unsigned int rq_nr_resend
;
1332 enum rq_phase rq_phase
; /* one of RQ_PHASE_* */
1333 enum rq_phase rq_next_phase
; /* one of RQ_PHASE_* to be used next */
1334 atomic_t rq_refcount
; /* client-side refcount for SENT race,
1335 * server-side refcount for multiple replies
1338 /** Portal to which this request would be sent */
1339 short rq_request_portal
; /* XXX FIXME bug 249 */
1340 /** Portal where to wait for reply and where reply would be sent */
1341 short rq_reply_portal
; /* XXX FIXME bug 249 */
1345 * !rq_truncate : # reply bytes actually received,
1346 * rq_truncate : required repbuf_len for resend
1348 int rq_nob_received
;
1349 /** Request length */
1353 /** Request message - what client sent */
1354 struct lustre_msg
*rq_reqmsg
;
1355 /** Reply message - server response */
1356 struct lustre_msg
*rq_repmsg
;
1357 /** Transaction number */
1362 * List item to for replay list. Not yet committed requests get linked
1364 * Also see \a rq_replay comment above.
1366 struct list_head rq_replay_list
;
1369 * security and encryption data
1372 struct ptlrpc_cli_ctx
*rq_cli_ctx
; /**< client's half ctx */
1373 struct ptlrpc_svc_ctx
*rq_svc_ctx
; /**< server's half ctx */
1374 struct list_head rq_ctx_chain
; /**< link to waited ctx */
1376 struct sptlrpc_flavor rq_flvr
; /**< for client & server */
1377 enum lustre_sec_part rq_sp_from
;
1379 /* client/server security flags */
1381 rq_ctx_init
:1, /* context initiation */
1382 rq_ctx_fini
:1, /* context destroy */
1383 rq_bulk_read
:1, /* request bulk read */
1384 rq_bulk_write
:1, /* request bulk write */
1385 /* server authentication flags */
1386 rq_auth_gss
:1, /* authenticated by gss */
1387 rq_auth_remote
:1, /* authed as remote user */
1388 rq_auth_usr_root
:1, /* authed as root */
1389 rq_auth_usr_mdt
:1, /* authed as mdt */
1390 rq_auth_usr_ost
:1, /* authed as ost */
1391 /* security tfm flags */
1394 /* doesn't expect reply FIXME */
1396 rq_pill_init
:1; /* pill initialized */
1398 uid_t rq_auth_uid
; /* authed uid */
1399 uid_t rq_auth_mapped_uid
; /* authed uid mapped to */
1401 /* (server side), pointed directly into req buffer */
1402 struct ptlrpc_user_desc
*rq_user_desc
;
1404 /* various buffer pointers */
1405 struct lustre_msg
*rq_reqbuf
; /* req wrapper */
1406 char *rq_repbuf
; /* rep buffer */
1407 struct lustre_msg
*rq_repdata
; /* rep wrapper msg */
1408 struct lustre_msg
*rq_clrbuf
; /* only in priv mode */
1409 int rq_reqbuf_len
; /* req wrapper buf len */
1410 int rq_reqdata_len
; /* req wrapper msg len */
1411 int rq_repbuf_len
; /* rep buffer len */
1412 int rq_repdata_len
; /* rep wrapper msg len */
1413 int rq_clrbuf_len
; /* only in priv mode */
1414 int rq_clrdata_len
; /* only in priv mode */
1416 /** early replies go to offset 0, regular replies go after that */
1417 unsigned int rq_reply_off
;
1421 /** Fields that help to see if request and reply were swabbed or not */
1422 __u32 rq_req_swab_mask
;
1423 __u32 rq_rep_swab_mask
;
1425 /** What was import generation when this request was sent */
1426 int rq_import_generation
;
1427 enum lustre_imp_state rq_send_state
;
1429 /** how many early replies (for stats) */
1432 /** client+server request */
1433 lnet_handle_md_t rq_req_md_h
;
1434 struct ptlrpc_cb_id rq_req_cbid
;
1435 /** optional time limit for send attempts */
1436 long rq_delay_limit
;
1437 /** time request was first queued */
1438 unsigned long rq_queued_time
;
1440 /* server-side... */
1441 /** request arrival time */
1442 struct timespec64 rq_arrival_time
;
1443 /** separated reply state */
1444 struct ptlrpc_reply_state
*rq_reply_state
;
1445 /** incoming request buffer */
1446 struct ptlrpc_request_buffer_desc
*rq_rqbd
;
1448 /** client-only incoming reply */
1449 lnet_handle_md_t rq_reply_md_h
;
1450 wait_queue_head_t rq_reply_waitq
;
1451 struct ptlrpc_cb_id rq_reply_cbid
;
1455 /** Peer description (the other side) */
1456 lnet_process_id_t rq_peer
;
1457 /** Server-side, export on which request was received */
1458 struct obd_export
*rq_export
;
1459 /** Client side, import where request is being sent */
1460 struct obd_import
*rq_import
;
1462 /** Replay callback, called after request is replayed at recovery */
1463 void (*rq_replay_cb
)(struct ptlrpc_request
*);
1465 * Commit callback, called when request is committed and about to be
1468 void (*rq_commit_cb
)(struct ptlrpc_request
*);
1469 /** Opaq data for replay and commit callbacks. */
1472 /** For bulk requests on client only: bulk descriptor */
1473 struct ptlrpc_bulk_desc
*rq_bulk
;
1475 /** client outgoing req */
1477 * when request/reply sent (secs), or time when request should be sent
1480 /** time for request really sent out */
1481 time64_t rq_real_sent
;
1483 /** when request must finish. volatile
1484 * so that servers' early reply updates to the deadline aren't
1485 * kept in per-cpu cache
1487 volatile time64_t rq_deadline
;
1488 /** when req reply unlink must finish. */
1489 time64_t rq_reply_deadline
;
1490 /** when req bulk unlink must finish. */
1491 time64_t rq_bulk_deadline
;
1493 * service time estimate (secs)
1494 * If the requestsis not served by this time, it is marked as timed out.
1498 /** Multi-rpc bits */
1499 /** Per-request waitq introduced by bug 21938 for recovery waiting */
1500 wait_queue_head_t rq_set_waitq
;
1501 /** Link item for request set lists */
1502 struct list_head rq_set_chain
;
1503 /** Link back to the request set */
1504 struct ptlrpc_request_set
*rq_set
;
1505 /** Async completion handler, called when reply is received */
1506 ptlrpc_interpterer_t rq_interpret_reply
;
1507 /** Async completion context */
1508 union ptlrpc_async_args rq_async_args
;
1510 /** Pool if request is from preallocated list */
1511 struct ptlrpc_request_pool
*rq_pool
;
1513 struct lu_context rq_session
;
1514 struct lu_context rq_recov_session
;
1516 /** request format description */
1517 struct req_capsule rq_pill
;
1521 * Call completion handler for rpc if any, return it's status or original
1522 * rc if there was no handler defined for this request.
1524 static inline int ptlrpc_req_interpret(const struct lu_env
*env
,
1525 struct ptlrpc_request
*req
, int rc
)
1527 if (req
->rq_interpret_reply
) {
1528 req
->rq_status
= req
->rq_interpret_reply(env
, req
,
1529 &req
->rq_async_args
,
1531 return req
->rq_status
;
1537 * Can the request be moved from the regular NRS head to the high-priority NRS
1538 * head (of the same PTLRPC service partition), if any?
1540 * For a reliable result, this should be checked under svcpt->scp_req lock.
1542 static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request
*req
)
1544 struct ptlrpc_nrs_request
*nrq
= &req
->rq_nrq
;
1547 * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the
1548 * request has been enqueued first, and ptlrpc_nrs_request::nr_started
1549 * to make sure it has not been scheduled yet (analogous to previous
1550 * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list).
1552 return nrq
->nr_enqueued
&& !nrq
->nr_started
&& !req
->rq_hp
;
1558 * Returns 1 if request buffer at offset \a index was already swabbed
1560 static inline int lustre_req_swabbed(struct ptlrpc_request
*req
, int index
)
1562 LASSERT(index
< sizeof(req
->rq_req_swab_mask
) * 8);
1563 return req
->rq_req_swab_mask
& (1 << index
);
1567 * Returns 1 if request reply buffer at offset \a index was already swabbed
1569 static inline int lustre_rep_swabbed(struct ptlrpc_request
*req
, int index
)
1571 LASSERT(index
< sizeof(req
->rq_rep_swab_mask
) * 8);
1572 return req
->rq_rep_swab_mask
& (1 << index
);
1576 * Returns 1 if request needs to be swabbed into local cpu byteorder
1578 static inline int ptlrpc_req_need_swab(struct ptlrpc_request
*req
)
1580 return lustre_req_swabbed(req
, MSG_PTLRPC_HEADER_OFF
);
1584 * Returns 1 if request reply needs to be swabbed into local cpu byteorder
1586 static inline int ptlrpc_rep_need_swab(struct ptlrpc_request
*req
)
1588 return lustre_rep_swabbed(req
, MSG_PTLRPC_HEADER_OFF
);
1592 * Mark request buffer at offset \a index that it was already swabbed
1594 static inline void lustre_set_req_swabbed(struct ptlrpc_request
*req
, int index
)
1596 LASSERT(index
< sizeof(req
->rq_req_swab_mask
) * 8);
1597 LASSERT((req
->rq_req_swab_mask
& (1 << index
)) == 0);
1598 req
->rq_req_swab_mask
|= 1 << index
;
1602 * Mark request reply buffer at offset \a index that it was already swabbed
1604 static inline void lustre_set_rep_swabbed(struct ptlrpc_request
*req
, int index
)
1606 LASSERT(index
< sizeof(req
->rq_rep_swab_mask
) * 8);
1607 LASSERT((req
->rq_rep_swab_mask
& (1 << index
)) == 0);
1608 req
->rq_rep_swab_mask
|= 1 << index
;
1612 * Convert numerical request phase value \a phase into text string description
1614 static inline const char *
1615 ptlrpc_phase2str(enum rq_phase phase
)
1624 case RQ_PHASE_INTERPRET
:
1626 case RQ_PHASE_COMPLETE
:
1628 case RQ_PHASE_UNREGISTERING
:
1629 return "Unregistering";
1636 * Convert numerical request phase of the request \a req into text stringi
1639 static inline const char *
1640 ptlrpc_rqphase2str(struct ptlrpc_request
*req
)
1642 return ptlrpc_phase2str(req
->rq_phase
);
1646 * Debugging functions and helpers to print request structure into debug log
1649 /* Spare the preprocessor, spoil the bugs. */
1650 #define FLAG(field, str) (field ? str : "")
1652 /** Convert bit flags into a string */
1653 #define DEBUG_REQ_FLAGS(req) \
1654 ptlrpc_rqphase2str(req), \
1655 FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \
1656 FLAG(req->rq_err, "E"), \
1657 FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \
1658 FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \
1659 FLAG(req->rq_no_resend, "N"), \
1660 FLAG(req->rq_waiting, "W"), \
1661 FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \
1662 FLAG(req->rq_committed, "M")
1664 #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s"
1666 void _debug_req(struct ptlrpc_request
*req
,
1667 struct libcfs_debug_msg_data
*data
, const char *fmt
, ...)
1671 * Helper that decides if we need to print request according to current debug
1674 #define debug_req(msgdata, mask, cdls, req, fmt, a...) \
1676 CFS_CHECK_STACK(msgdata, mask, cdls); \
1678 if (((mask) & D_CANTMASK) != 0 || \
1679 ((libcfs_debug & (mask)) != 0 && \
1680 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1681 _debug_req((req), msgdata, fmt, ##a); \
1685 * This is the debug print function you need to use to print request structure
1686 * content into lustre debug log.
1687 * for most callers (level is a constant) this is resolved at compile time
1689 #define DEBUG_REQ(level, req, fmt, args...) \
1691 if ((level) & (D_ERROR | D_WARNING)) { \
1692 static struct cfs_debug_limit_state cdls; \
1693 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1694 debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\
1696 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1697 debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \
1703 * Structure that defines a single page of a bulk transfer
1705 struct ptlrpc_bulk_page
{
1706 /** Linkage to list of pages in a bulk */
1707 struct list_head bp_link
;
1709 * Number of bytes in a page to transfer starting from \a bp_pageoffset
1712 /** offset within a page */
1714 /** The page itself */
1715 struct page
*bp_page
;
1718 #define BULK_GET_SOURCE 0
1719 #define BULK_PUT_SINK 1
1720 #define BULK_GET_SINK 2
1721 #define BULK_PUT_SOURCE 3
1724 * Definition of bulk descriptor.
1725 * Bulks are special "Two phase" RPCs where initial request message
1726 * is sent first and it is followed bt a transfer (o receiving) of a large
1727 * amount of data to be settled into pages referenced from the bulk descriptors.
1728 * Bulks transfers (the actual data following the small requests) are done
1729 * on separate LNet portals.
1730 * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs.
1731 * Another user is readpage for MDT.
1733 struct ptlrpc_bulk_desc
{
1734 /** completed with failure */
1735 unsigned long bd_failure
:1;
1736 /** {put,get}{source,sink} */
1737 unsigned long bd_type
:2;
1739 unsigned long bd_registered
:1;
1740 /** For serialization with callback */
1742 /** Import generation when request for this bulk was sent */
1743 int bd_import_generation
;
1744 /** LNet portal for this bulk */
1746 /** Server side - export this bulk created for */
1747 struct obd_export
*bd_export
;
1748 /** Client side - import this bulk was sent on */
1749 struct obd_import
*bd_import
;
1750 /** Back pointer to the request */
1751 struct ptlrpc_request
*bd_req
;
1752 wait_queue_head_t bd_waitq
; /* server side only WQ */
1753 int bd_iov_count
; /* # entries in bd_iov */
1754 int bd_max_iov
; /* allocated size of bd_iov */
1755 int bd_nob
; /* # bytes covered */
1756 int bd_nob_transferred
; /* # bytes GOT/PUT */
1760 struct ptlrpc_cb_id bd_cbid
; /* network callback info */
1761 lnet_nid_t bd_sender
; /* stash event::sender */
1762 int bd_md_count
; /* # valid entries in bd_mds */
1763 int bd_md_max_brw
; /* max entries in bd_mds */
1764 /** array of associated MDs */
1765 lnet_handle_md_t bd_mds
[PTLRPC_BULK_OPS_COUNT
];
1768 * encrypt iov, size is either 0 or bd_iov_count.
1770 lnet_kiov_t
*bd_enc_iov
;
1772 lnet_kiov_t bd_iov
[0];
1776 SVC_STOPPED
= 1 << 0,
1777 SVC_STOPPING
= 1 << 1,
1778 SVC_STARTING
= 1 << 2,
1779 SVC_RUNNING
= 1 << 3,
1781 SVC_SIGNAL
= 1 << 5,
1784 #define PTLRPC_THR_NAME_LEN 32
1786 * Definition of server service thread structure
1788 struct ptlrpc_thread
{
1790 * List of active threads in svc->srv_threads
1792 struct list_head t_link
;
1794 * thread-private data (preallocated memory)
1799 * service thread index, from ptlrpc_start_threads
1803 * service thread pid
1807 * put watchdog in the structure per thread b=14840
1809 * Lustre watchdog is removed for client in the hope
1810 * of a generic watchdog can be merged in kernel.
1811 * When that happens, we should add below back.
1813 * struct lc_watchdog *t_watchdog;
1816 * the svc this thread belonged to b=18582
1818 struct ptlrpc_service_part
*t_svcpt
;
1819 wait_queue_head_t t_ctl_waitq
;
1820 struct lu_env
*t_env
;
1821 char t_name
[PTLRPC_THR_NAME_LEN
];
1824 static inline int thread_is_init(struct ptlrpc_thread
*thread
)
1826 return thread
->t_flags
== 0;
1829 static inline int thread_is_stopped(struct ptlrpc_thread
*thread
)
1831 return !!(thread
->t_flags
& SVC_STOPPED
);
1834 static inline int thread_is_stopping(struct ptlrpc_thread
*thread
)
1836 return !!(thread
->t_flags
& SVC_STOPPING
);
1839 static inline int thread_is_starting(struct ptlrpc_thread
*thread
)
1841 return !!(thread
->t_flags
& SVC_STARTING
);
1844 static inline int thread_is_running(struct ptlrpc_thread
*thread
)
1846 return !!(thread
->t_flags
& SVC_RUNNING
);
1849 static inline int thread_is_event(struct ptlrpc_thread
*thread
)
1851 return !!(thread
->t_flags
& SVC_EVENT
);
1854 static inline int thread_is_signal(struct ptlrpc_thread
*thread
)
1856 return !!(thread
->t_flags
& SVC_SIGNAL
);
1859 static inline void thread_clear_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
1861 thread
->t_flags
&= ~flags
;
1864 static inline void thread_set_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
1866 thread
->t_flags
= flags
;
1869 static inline void thread_add_flags(struct ptlrpc_thread
*thread
, __u32 flags
)
1871 thread
->t_flags
|= flags
;
1874 static inline int thread_test_and_clear_flags(struct ptlrpc_thread
*thread
,
1877 if (thread
->t_flags
& flags
) {
1878 thread
->t_flags
&= ~flags
;
1885 * Request buffer descriptor structure.
1886 * This is a structure that contains one posted request buffer for service.
1887 * Once data land into a buffer, event callback creates actual request and
1888 * notifies wakes one of the service threads to process new incoming request.
1889 * More than one request can fit into the buffer.
1891 struct ptlrpc_request_buffer_desc
{
1892 /** Link item for rqbds on a service */
1893 struct list_head rqbd_list
;
1894 /** History of requests for this buffer */
1895 struct list_head rqbd_reqs
;
1896 /** Back pointer to service for which this buffer is registered */
1897 struct ptlrpc_service_part
*rqbd_svcpt
;
1898 /** LNet descriptor */
1899 lnet_handle_md_t rqbd_md_h
;
1901 /** The buffer itself */
1903 struct ptlrpc_cb_id rqbd_cbid
;
1905 * This "embedded" request structure is only used for the
1906 * last request to fit into the buffer
1908 struct ptlrpc_request rqbd_req
;
1911 typedef int (*svc_handler_t
)(struct ptlrpc_request
*req
);
1913 struct ptlrpc_service_ops
{
1915 * if non-NULL called during thread creation (ptlrpc_start_thread())
1916 * to initialize service specific per-thread state.
1918 int (*so_thr_init
)(struct ptlrpc_thread
*thr
);
1920 * if non-NULL called during thread shutdown (ptlrpc_main()) to
1921 * destruct state created by ->srv_init().
1923 void (*so_thr_done
)(struct ptlrpc_thread
*thr
);
1925 * Handler function for incoming requests for this service
1927 int (*so_req_handler
)(struct ptlrpc_request
*req
);
1929 * function to determine priority of the request, it's called
1930 * on every new request
1932 int (*so_hpreq_handler
)(struct ptlrpc_request
*);
1934 * service-specific print fn
1936 void (*so_req_printer
)(void *, struct ptlrpc_request
*);
1939 #ifndef __cfs_cacheline_aligned
1940 /* NB: put it here for reducing patche dependence */
1941 # define __cfs_cacheline_aligned
1945 * How many high priority requests to serve before serving one normal
1948 #define PTLRPC_SVC_HP_RATIO 10
1951 * Definition of PortalRPC service.
1952 * The service is listening on a particular portal (like tcp port)
1953 * and perform actions for a specific server like IO service for OST
1954 * or general metadata service for MDS.
1956 struct ptlrpc_service
{
1957 /** serialize sysfs operations */
1958 spinlock_t srv_lock
;
1959 /** most often accessed fields */
1960 /** chain thru all services */
1961 struct list_head srv_list
;
1962 /** service operations table */
1963 struct ptlrpc_service_ops srv_ops
;
1964 /** only statically allocated strings here; we don't clean them */
1966 /** only statically allocated strings here; we don't clean them */
1967 char *srv_thread_name
;
1968 /** service thread list */
1969 struct list_head srv_threads
;
1970 /** threads # should be created for each partition on initializing */
1971 int srv_nthrs_cpt_init
;
1972 /** limit of threads number for each partition */
1973 int srv_nthrs_cpt_limit
;
1974 /** Root of debugfs dir tree for this service */
1975 struct dentry
*srv_debugfs_entry
;
1976 /** Pointer to statistic data for this service */
1977 struct lprocfs_stats
*srv_stats
;
1978 /** # hp per lp reqs to handle */
1979 int srv_hpreq_ratio
;
1980 /** biggest request to receive */
1981 int srv_max_req_size
;
1982 /** biggest reply to send */
1983 int srv_max_reply_size
;
1984 /** size of individual buffers */
1986 /** # buffers to allocate in 1 group */
1987 int srv_nbuf_per_group
;
1988 /** Local portal on which to receive requests */
1989 __u32 srv_req_portal
;
1990 /** Portal on the client to send replies to */
1991 __u32 srv_rep_portal
;
1993 * Tags for lu_context associated with this thread, see struct
1997 /** soft watchdog timeout multiplier */
1998 int srv_watchdog_factor
;
1999 /** under unregister_service */
2000 unsigned srv_is_stopping
:1;
2002 /** max # request buffers in history per partition */
2003 int srv_hist_nrqbds_cpt_max
;
2004 /** number of CPTs this service bound on */
2006 /** CPTs array this service bound on */
2008 /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */
2010 /** CPT table this service is running over */
2011 struct cfs_cpt_table
*srv_cptable
;
2014 struct kobject srv_kobj
;
2015 struct completion srv_kobj_unregister
;
2017 * partition data for ptlrpc service
2019 struct ptlrpc_service_part
*srv_parts
[0];
2023 * Definition of PortalRPC service partition data.
2024 * Although a service only has one instance of it right now, but we
2025 * will have multiple instances very soon (instance per CPT).
2027 * it has four locks:
2029 * serialize operations on rqbd and requests waiting for preprocess
2031 * serialize operations active requests sent to this portal
2033 * serialize adaptive timeout stuff
2035 * serialize operations on RS list (reply states)
2037 * We don't have any use-case to take two or more locks at the same time
2038 * for now, so there is no lock order issue.
2040 struct ptlrpc_service_part
{
2041 /** back reference to owner */
2042 struct ptlrpc_service
*scp_service __cfs_cacheline_aligned
;
2043 /* CPT id, reserved */
2045 /** always increasing number */
2047 /** # of starting threads */
2048 int scp_nthrs_starting
;
2049 /** # of stopping threads, reserved for shrinking threads */
2050 int scp_nthrs_stopping
;
2051 /** # running threads */
2052 int scp_nthrs_running
;
2053 /** service threads list */
2054 struct list_head scp_threads
;
2057 * serialize the following fields, used for protecting
2058 * rqbd list and incoming requests waiting for preprocess,
2059 * threads starting & stopping are also protected by this lock.
2061 spinlock_t scp_lock __cfs_cacheline_aligned
;
2062 /** total # req buffer descs allocated */
2063 int scp_nrqbds_total
;
2064 /** # posted request buffers for receiving */
2065 int scp_nrqbds_posted
;
2066 /** in progress of allocating rqbd */
2067 int scp_rqbd_allocating
;
2068 /** # incoming reqs */
2069 int scp_nreqs_incoming
;
2070 /** request buffers to be reposted */
2071 struct list_head scp_rqbd_idle
;
2072 /** req buffers receiving */
2073 struct list_head scp_rqbd_posted
;
2074 /** incoming reqs */
2075 struct list_head scp_req_incoming
;
2076 /** timeout before re-posting reqs, in tick */
2077 long scp_rqbd_timeout
;
2079 * all threads sleep on this. This wait-queue is signalled when new
2080 * incoming request arrives and when difficult reply has to be handled.
2082 wait_queue_head_t scp_waitq
;
2084 /** request history */
2085 struct list_head scp_hist_reqs
;
2086 /** request buffer history */
2087 struct list_head scp_hist_rqbds
;
2088 /** # request buffers in history */
2089 int scp_hist_nrqbds
;
2090 /** sequence number for request */
2092 /** highest seq culled from history */
2093 __u64 scp_hist_seq_culled
;
2096 * serialize the following fields, used for processing requests
2097 * sent to this portal
2099 spinlock_t scp_req_lock __cfs_cacheline_aligned
;
2100 /** # reqs in either of the NRS heads below */
2101 /** # reqs being served */
2102 int scp_nreqs_active
;
2103 /** # HPreqs being served */
2104 int scp_nhreqs_active
;
2105 /** # hp requests handled */
2108 /** NRS head for regular requests */
2109 struct ptlrpc_nrs scp_nrs_reg
;
2110 /** NRS head for HP requests; this is only valid for services that can
2111 * handle HP requests
2113 struct ptlrpc_nrs
*scp_nrs_hp
;
2118 * serialize the following fields, used for changes on
2121 spinlock_t scp_at_lock __cfs_cacheline_aligned
;
2122 /** estimated rpc service time */
2123 struct adaptive_timeout scp_at_estimate
;
2124 /** reqs waiting for replies */
2125 struct ptlrpc_at_array scp_at_array
;
2126 /** early reply timer */
2127 struct timer_list scp_at_timer
;
2129 unsigned long scp_at_checktime
;
2130 /** check early replies */
2131 unsigned scp_at_check
;
2135 * serialize the following fields, used for processing
2136 * replies for this portal
2138 spinlock_t scp_rep_lock __cfs_cacheline_aligned
;
2139 /** all the active replies */
2140 struct list_head scp_rep_active
;
2141 /** List of free reply_states */
2142 struct list_head scp_rep_idle
;
2143 /** waitq to run, when adding stuff to srv_free_rs_list */
2144 wait_queue_head_t scp_rep_waitq
;
2145 /** # 'difficult' replies */
2146 atomic_t scp_nreps_difficult
;
2149 #define ptlrpc_service_for_each_part(part, i, svc) \
2151 i < (svc)->srv_ncpts && \
2152 (svc)->srv_parts && \
2153 ((part) = (svc)->srv_parts[i]); i++)
2156 * Declaration of ptlrpcd control structure
2158 struct ptlrpcd_ctl
{
2160 * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE)
2162 unsigned long pc_flags
;
2164 * Thread lock protecting structure fields.
2170 struct completion pc_starting
;
2174 struct completion pc_finishing
;
2176 * Thread requests set.
2178 struct ptlrpc_request_set
*pc_set
;
2180 * Thread name used in kthread_run()
2184 * Environment for request interpreters to run in.
2186 struct lu_env pc_env
;
2188 * CPT the thread is bound on.
2192 * Index of ptlrpcd thread in the array.
2196 * Pointer to the array of partners' ptlrpcd_ctl structure.
2198 struct ptlrpcd_ctl
**pc_partners
;
2200 * Number of the ptlrpcd's partners.
2204 * Record the partner index to be processed next.
2208 * Error code if the thread failed to fully start.
2213 /* Bits for pc_flags */
2214 enum ptlrpcd_ctl_flags
{
2216 * Ptlrpc thread start flag.
2218 LIOD_START
= 1 << 0,
2220 * Ptlrpc thread stop flag.
2224 * Ptlrpc thread force flag (only stop force so far).
2225 * This will cause aborting any inflight rpcs handled
2226 * by thread if LIOD_STOP is specified.
2228 LIOD_FORCE
= 1 << 2,
2230 * This is a recovery ptlrpc thread.
2232 LIOD_RECOVERY
= 1 << 3,
2239 * Service compatibility function; the policy is compatible with all services.
2241 * \param[in] svc The service the policy is attempting to register with.
2242 * \param[in] desc The policy descriptor
2244 * \retval true The policy is compatible with the service
2246 * \see ptlrpc_nrs_pol_desc::pd_compat()
2248 static inline bool nrs_policy_compat_all(const struct ptlrpc_service
*svc
,
2249 const struct ptlrpc_nrs_pol_desc
*desc
)
2255 * Service compatibility function; the policy is compatible with only a specific
2256 * service which is identified by its human-readable name at
2257 * ptlrpc_service::srv_name.
2259 * \param[in] svc The service the policy is attempting to register with.
2260 * \param[in] desc The policy descriptor
2262 * \retval false The policy is not compatible with the service
2263 * \retval true The policy is compatible with the service
2265 * \see ptlrpc_nrs_pol_desc::pd_compat()
2267 static inline bool nrs_policy_compat_one(const struct ptlrpc_service
*svc
,
2268 const struct ptlrpc_nrs_pol_desc
*desc
)
2270 return strcmp(svc
->srv_name
, desc
->pd_compat_svc_name
) == 0;
2275 /* ptlrpc/events.c */
2276 extern lnet_handle_eq_t ptlrpc_eq_h
;
2277 int ptlrpc_uuid_to_peer(struct obd_uuid
*uuid
,
2278 lnet_process_id_t
*peer
, lnet_nid_t
*self
);
2280 * These callbacks are invoked by LNet when something happened to
2284 void request_out_callback(lnet_event_t
*ev
);
2285 void reply_in_callback(lnet_event_t
*ev
);
2286 void client_bulk_callback(lnet_event_t
*ev
);
2287 void request_in_callback(lnet_event_t
*ev
);
2288 void reply_out_callback(lnet_event_t
*ev
);
2291 /* ptlrpc/connection.c */
2292 struct ptlrpc_connection
*ptlrpc_connection_get(lnet_process_id_t peer
,
2294 struct obd_uuid
*uuid
);
2295 int ptlrpc_connection_put(struct ptlrpc_connection
*c
);
2296 struct ptlrpc_connection
*ptlrpc_connection_addref(struct ptlrpc_connection
*);
2297 int ptlrpc_connection_init(void);
2298 void ptlrpc_connection_fini(void);
2300 /* ptlrpc/niobuf.c */
2302 * Actual interfacing with LNet to put/get/register/unregister stuff
2306 int ptlrpc_unregister_bulk(struct ptlrpc_request
*req
, int async
);
2308 static inline int ptlrpc_client_bulk_active(struct ptlrpc_request
*req
)
2310 struct ptlrpc_bulk_desc
*desc
;
2313 desc
= req
->rq_bulk
;
2315 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK
) &&
2316 req
->rq_bulk_deadline
> ktime_get_real_seconds())
2322 spin_lock(&desc
->bd_lock
);
2323 rc
= desc
->bd_md_count
;
2324 spin_unlock(&desc
->bd_lock
);
2328 #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01
2329 #define PTLRPC_REPLY_EARLY 0x02
2330 int ptlrpc_send_reply(struct ptlrpc_request
*req
, int flags
);
2331 int ptlrpc_reply(struct ptlrpc_request
*req
);
2332 int ptlrpc_send_error(struct ptlrpc_request
*req
, int difficult
);
2333 int ptlrpc_error(struct ptlrpc_request
*req
);
2334 void ptlrpc_resend_req(struct ptlrpc_request
*request
);
2335 int ptlrpc_at_get_net_latency(struct ptlrpc_request
*req
);
2336 int ptl_send_rpc(struct ptlrpc_request
*request
, int noreply
);
2337 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc
*rqbd
);
2340 /* ptlrpc/client.c */
2342 * Client-side portals API. Everything to send requests, receive replies,
2343 * request queues, request management, etc.
2346 void ptlrpc_request_committed(struct ptlrpc_request
*req
, int force
);
2348 void ptlrpc_init_client(int req_portal
, int rep_portal
, char *name
,
2349 struct ptlrpc_client
*);
2350 struct ptlrpc_connection
*ptlrpc_uuid_to_connection(struct obd_uuid
*uuid
);
2352 int ptlrpc_queue_wait(struct ptlrpc_request
*req
);
2353 int ptlrpc_replay_req(struct ptlrpc_request
*req
);
2354 int ptlrpc_unregister_reply(struct ptlrpc_request
*req
, int async
);
2355 void ptlrpc_abort_inflight(struct obd_import
*imp
);
2356 void ptlrpc_abort_set(struct ptlrpc_request_set
*set
);
2358 struct ptlrpc_request_set
*ptlrpc_prep_set(void);
2359 struct ptlrpc_request_set
*ptlrpc_prep_fcset(int max
, set_producer_func func
,
2361 int ptlrpc_set_next_timeout(struct ptlrpc_request_set
*);
2362 int ptlrpc_check_set(const struct lu_env
*env
, struct ptlrpc_request_set
*set
);
2363 int ptlrpc_set_wait(struct ptlrpc_request_set
*);
2364 int ptlrpc_expired_set(void *data
);
2365 void ptlrpc_interrupted_set(void *data
);
2366 void ptlrpc_mark_interrupted(struct ptlrpc_request
*req
);
2367 void ptlrpc_set_destroy(struct ptlrpc_request_set
*);
2368 void ptlrpc_set_add_req(struct ptlrpc_request_set
*, struct ptlrpc_request
*);
2369 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl
*pc
,
2370 struct ptlrpc_request
*req
);
2372 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool
*pool
);
2373 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool
*pool
, int num_rq
);
2375 struct ptlrpc_request_pool
*
2376 ptlrpc_init_rq_pool(int, int,
2377 int (*populate_pool
)(struct ptlrpc_request_pool
*, int));
2379 void ptlrpc_at_set_req_timeout(struct ptlrpc_request
*req
);
2380 struct ptlrpc_request
*ptlrpc_request_alloc(struct obd_import
*imp
,
2381 const struct req_format
*format
);
2382 struct ptlrpc_request
*ptlrpc_request_alloc_pool(struct obd_import
*imp
,
2383 struct ptlrpc_request_pool
*,
2384 const struct req_format
*);
2385 void ptlrpc_request_free(struct ptlrpc_request
*request
);
2386 int ptlrpc_request_pack(struct ptlrpc_request
*request
,
2387 __u32 version
, int opcode
);
2388 struct ptlrpc_request
*ptlrpc_request_alloc_pack(struct obd_import
*,
2389 const struct req_format
*,
2391 int ptlrpc_request_bufs_pack(struct ptlrpc_request
*request
,
2392 __u32 version
, int opcode
, char **bufs
,
2393 struct ptlrpc_cli_ctx
*ctx
);
2394 void ptlrpc_req_finished(struct ptlrpc_request
*request
);
2395 struct ptlrpc_request
*ptlrpc_request_addref(struct ptlrpc_request
*req
);
2396 struct ptlrpc_bulk_desc
*ptlrpc_prep_bulk_imp(struct ptlrpc_request
*req
,
2397 unsigned npages
, unsigned max_brw
,
2398 unsigned type
, unsigned portal
);
2399 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc
*bulk
, int pin
);
2400 static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc
*bulk
)
2402 __ptlrpc_free_bulk(bulk
, 1);
2405 static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc
*bulk
)
2407 __ptlrpc_free_bulk(bulk
, 0);
2410 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc
*desc
,
2411 struct page
*page
, int pageoffset
, int len
, int);
2412 static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc
*desc
,
2413 struct page
*page
, int pageoffset
,
2416 __ptlrpc_prep_bulk_page(desc
, page
, pageoffset
, len
, 1);
2419 static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc
*desc
,
2420 struct page
*page
, int pageoffset
,
2423 __ptlrpc_prep_bulk_page(desc
, page
, pageoffset
, len
, 0);
2426 void ptlrpc_retain_replayable_request(struct ptlrpc_request
*req
,
2427 struct obd_import
*imp
);
2428 __u64
ptlrpc_next_xid(void);
2429 __u64
ptlrpc_sample_next_xid(void);
2430 __u64
ptlrpc_req_xid(struct ptlrpc_request
*request
);
2432 /* Set of routines to run a function in ptlrpcd context */
2433 void *ptlrpcd_alloc_work(struct obd_import
*imp
,
2434 int (*cb
)(const struct lu_env
*, void *), void *data
);
2435 void ptlrpcd_destroy_work(void *handler
);
2436 int ptlrpcd_queue_work(void *handler
);
2439 struct ptlrpc_service_buf_conf
{
2440 /* nbufs is buffers # to allocate when growing the pool */
2441 unsigned int bc_nbufs
;
2442 /* buffer size to post */
2443 unsigned int bc_buf_size
;
2444 /* portal to listed for requests on */
2445 unsigned int bc_req_portal
;
2446 /* portal of where to send replies to */
2447 unsigned int bc_rep_portal
;
2448 /* maximum request size to be accepted for this service */
2449 unsigned int bc_req_max_size
;
2450 /* maximum reply size this service can ever send */
2451 unsigned int bc_rep_max_size
;
2454 struct ptlrpc_service_thr_conf
{
2455 /* threadname should be 8 characters or less - 6 will be added on */
2457 /* threads increasing factor for each CPU */
2458 unsigned int tc_thr_factor
;
2459 /* service threads # to start on each partition while initializing */
2460 unsigned int tc_nthrs_init
;
2462 * low water of threads # upper-limit on each partition while running,
2463 * service availability may be impacted if threads number is lower
2464 * than this value. It can be ZERO if the service doesn't require
2465 * CPU affinity or there is only one partition.
2467 unsigned int tc_nthrs_base
;
2468 /* "soft" limit for total threads number */
2469 unsigned int tc_nthrs_max
;
2470 /* user specified threads number, it will be validated due to
2471 * other members of this structure.
2473 unsigned int tc_nthrs_user
;
2474 /* set NUMA node affinity for service threads */
2475 unsigned int tc_cpu_affinity
;
2476 /* Tags for lu_context associated with service thread */
2480 struct ptlrpc_service_cpt_conf
{
2481 struct cfs_cpt_table
*cc_cptable
;
2482 /* string pattern to describe CPTs for a service */
2486 struct ptlrpc_service_conf
{
2489 /* soft watchdog timeout multiplifier to print stuck service traces */
2490 unsigned int psc_watchdog_factor
;
2491 /* buffer information */
2492 struct ptlrpc_service_buf_conf psc_buf
;
2493 /* thread information */
2494 struct ptlrpc_service_thr_conf psc_thr
;
2495 /* CPU partition information */
2496 struct ptlrpc_service_cpt_conf psc_cpt
;
2497 /* function table */
2498 struct ptlrpc_service_ops psc_ops
;
2501 /* ptlrpc/service.c */
2503 * Server-side services API. Register/unregister service, request state
2504 * management, service thread management
2508 void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state
*rs
);
2509 void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state
*rs
);
2510 struct ptlrpc_service
*ptlrpc_register_service(struct ptlrpc_service_conf
*conf
,
2511 struct kset
*parent
,
2512 struct dentry
*debugfs_entry
);
2514 int ptlrpc_start_threads(struct ptlrpc_service
*svc
);
2515 int ptlrpc_unregister_service(struct ptlrpc_service
*service
);
2517 int ptlrpc_hr_init(void);
2518 void ptlrpc_hr_fini(void);
2522 /* ptlrpc/import.c */
2527 int ptlrpc_connect_import(struct obd_import
*imp
);
2528 int ptlrpc_init_import(struct obd_import
*imp
);
2529 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
);
2530 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
);
2532 /* ptlrpc/pack_generic.c */
2533 int ptlrpc_reconnect_import(struct obd_import
*imp
);
2537 * ptlrpc msg buffer and swab interface
2541 int ptlrpc_buf_need_swab(struct ptlrpc_request
*req
, const int inout
,
2543 void ptlrpc_buf_set_swabbed(struct ptlrpc_request
*req
, const int inout
,
2545 int ptlrpc_unpack_rep_msg(struct ptlrpc_request
*req
, int len
);
2546 int ptlrpc_unpack_req_msg(struct ptlrpc_request
*req
, int len
);
2548 void lustre_init_msg_v2(struct lustre_msg_v2
*msg
, int count
, __u32
*lens
,
2550 int lustre_pack_request(struct ptlrpc_request
*, __u32 magic
, int count
,
2551 __u32
*lens
, char **bufs
);
2552 int lustre_pack_reply(struct ptlrpc_request
*, int count
, __u32
*lens
,
2554 int lustre_pack_reply_v2(struct ptlrpc_request
*req
, int count
,
2555 __u32
*lens
, char **bufs
, int flags
);
2556 #define LPRFL_EARLY_REPLY 1
2557 int lustre_pack_reply_flags(struct ptlrpc_request
*, int count
, __u32
*lens
,
2558 char **bufs
, int flags
);
2559 int lustre_shrink_msg(struct lustre_msg
*msg
, int segment
,
2560 unsigned int newlen
, int move_data
);
2561 void lustre_free_reply_state(struct ptlrpc_reply_state
*rs
);
2562 int __lustre_unpack_msg(struct lustre_msg
*m
, int len
);
2563 int lustre_msg_hdr_size(__u32 magic
, int count
);
2564 int lustre_msg_size(__u32 magic
, int count
, __u32
*lengths
);
2565 int lustre_msg_size_v2(int count
, __u32
*lengths
);
2566 int lustre_packed_msg_size(struct lustre_msg
*msg
);
2567 int lustre_msg_early_size(void);
2568 void *lustre_msg_buf_v2(struct lustre_msg_v2
*m
, int n
, int min_size
);
2569 void *lustre_msg_buf(struct lustre_msg
*m
, int n
, int minlen
);
2570 int lustre_msg_buflen(struct lustre_msg
*m
, int n
);
2571 int lustre_msg_bufcount(struct lustre_msg
*m
);
2572 char *lustre_msg_string(struct lustre_msg
*m
, int n
, int max_len
);
2573 __u32
lustre_msghdr_get_flags(struct lustre_msg
*msg
);
2574 void lustre_msghdr_set_flags(struct lustre_msg
*msg
, __u32 flags
);
2575 __u32
lustre_msg_get_flags(struct lustre_msg
*msg
);
2576 void lustre_msg_add_flags(struct lustre_msg
*msg
, int flags
);
2577 void lustre_msg_set_flags(struct lustre_msg
*msg
, int flags
);
2578 void lustre_msg_clear_flags(struct lustre_msg
*msg
, int flags
);
2579 __u32
lustre_msg_get_op_flags(struct lustre_msg
*msg
);
2580 void lustre_msg_add_op_flags(struct lustre_msg
*msg
, int flags
);
2581 struct lustre_handle
*lustre_msg_get_handle(struct lustre_msg
*msg
);
2582 __u32
lustre_msg_get_type(struct lustre_msg
*msg
);
2583 void lustre_msg_add_version(struct lustre_msg
*msg
, int version
);
2584 __u32
lustre_msg_get_opc(struct lustre_msg
*msg
);
2585 __u64
lustre_msg_get_last_committed(struct lustre_msg
*msg
);
2586 __u64
*lustre_msg_get_versions(struct lustre_msg
*msg
);
2587 __u64
lustre_msg_get_transno(struct lustre_msg
*msg
);
2588 __u64
lustre_msg_get_slv(struct lustre_msg
*msg
);
2589 __u32
lustre_msg_get_limit(struct lustre_msg
*msg
);
2590 void lustre_msg_set_slv(struct lustre_msg
*msg
, __u64 slv
);
2591 void lustre_msg_set_limit(struct lustre_msg
*msg
, __u64 limit
);
2592 int lustre_msg_get_status(struct lustre_msg
*msg
);
2593 __u32
lustre_msg_get_conn_cnt(struct lustre_msg
*msg
);
2594 __u32
lustre_msg_get_magic(struct lustre_msg
*msg
);
2595 __u32
lustre_msg_get_timeout(struct lustre_msg
*msg
);
2596 __u32
lustre_msg_get_service_time(struct lustre_msg
*msg
);
2597 __u32
lustre_msg_get_cksum(struct lustre_msg
*msg
);
2598 __u32
lustre_msg_calc_cksum(struct lustre_msg
*msg
);
2599 void lustre_msg_set_handle(struct lustre_msg
*msg
,
2600 struct lustre_handle
*handle
);
2601 void lustre_msg_set_type(struct lustre_msg
*msg
, __u32 type
);
2602 void lustre_msg_set_opc(struct lustre_msg
*msg
, __u32 opc
);
2603 void lustre_msg_set_versions(struct lustre_msg
*msg
, __u64
*versions
);
2604 void lustre_msg_set_transno(struct lustre_msg
*msg
, __u64 transno
);
2605 void lustre_msg_set_status(struct lustre_msg
*msg
, __u32 status
);
2606 void lustre_msg_set_conn_cnt(struct lustre_msg
*msg
, __u32 conn_cnt
);
2607 void ptlrpc_request_set_replen(struct ptlrpc_request
*req
);
2608 void lustre_msg_set_timeout(struct lustre_msg
*msg
, __u32 timeout
);
2609 void lustre_msg_set_service_time(struct lustre_msg
*msg
, __u32 service_time
);
2610 void lustre_msg_set_jobid(struct lustre_msg
*msg
, char *jobid
);
2611 void lustre_msg_set_cksum(struct lustre_msg
*msg
, __u32 cksum
);
2614 lustre_shrink_reply(struct ptlrpc_request
*req
, int segment
,
2615 unsigned int newlen
, int move_data
)
2617 LASSERT(req
->rq_reply_state
);
2618 LASSERT(req
->rq_repmsg
);
2619 req
->rq_replen
= lustre_shrink_msg(req
->rq_repmsg
, segment
,
2623 #ifdef CONFIG_LUSTRE_TRANSLATE_ERRNOS
2625 static inline int ptlrpc_status_hton(int h
)
2628 * Positive errnos must be network errnos, such as LUSTRE_EDEADLK,
2629 * ELDLM_LOCK_ABORTED, etc.
2632 return -lustre_errno_hton(-h
);
2637 static inline int ptlrpc_status_ntoh(int n
)
2640 * See the comment in ptlrpc_status_hton().
2643 return -lustre_errno_ntoh(-n
);
2650 #define ptlrpc_status_hton(h) (h)
2651 #define ptlrpc_status_ntoh(n) (n)
2656 /** Change request phase of \a req to \a new_phase */
2658 ptlrpc_rqphase_move(struct ptlrpc_request
*req
, enum rq_phase new_phase
)
2660 if (req
->rq_phase
== new_phase
)
2663 if (new_phase
== RQ_PHASE_UNREGISTERING
) {
2664 req
->rq_next_phase
= req
->rq_phase
;
2666 atomic_inc(&req
->rq_import
->imp_unregistering
);
2669 if (req
->rq_phase
== RQ_PHASE_UNREGISTERING
) {
2671 atomic_dec(&req
->rq_import
->imp_unregistering
);
2674 DEBUG_REQ(D_INFO
, req
, "move req \"%s\" -> \"%s\"",
2675 ptlrpc_rqphase2str(req
), ptlrpc_phase2str(new_phase
));
2677 req
->rq_phase
= new_phase
;
2681 * Returns true if request \a req got early reply and hard deadline is not met
2684 ptlrpc_client_early(struct ptlrpc_request
*req
)
2686 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2687 req
->rq_reply_deadline
> ktime_get_real_seconds())
2689 return req
->rq_early
;
2693 * Returns true if we got real reply from server for this request
2696 ptlrpc_client_replied(struct ptlrpc_request
*req
)
2698 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2699 req
->rq_reply_deadline
> ktime_get_real_seconds())
2701 return req
->rq_replied
;
2704 /** Returns true if request \a req is in process of receiving server reply */
2706 ptlrpc_client_recv(struct ptlrpc_request
*req
)
2708 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2709 req
->rq_reply_deadline
> ktime_get_real_seconds())
2711 return req
->rq_receiving_reply
;
2715 ptlrpc_client_recv_or_unlink(struct ptlrpc_request
*req
)
2719 spin_lock(&req
->rq_lock
);
2720 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
) &&
2721 req
->rq_reply_deadline
> ktime_get_real_seconds()) {
2722 spin_unlock(&req
->rq_lock
);
2725 rc
= req
->rq_receiving_reply
;
2726 rc
= rc
|| req
->rq_req_unlink
|| req
->rq_reply_unlink
;
2727 spin_unlock(&req
->rq_lock
);
2732 ptlrpc_client_wake_req(struct ptlrpc_request
*req
)
2735 wake_up(&req
->rq_reply_waitq
);
2737 wake_up(&req
->rq_set
->set_waitq
);
2741 ptlrpc_rs_addref(struct ptlrpc_reply_state
*rs
)
2743 LASSERT(atomic_read(&rs
->rs_refcount
) > 0);
2744 atomic_inc(&rs
->rs_refcount
);
2748 ptlrpc_rs_decref(struct ptlrpc_reply_state
*rs
)
2750 LASSERT(atomic_read(&rs
->rs_refcount
) > 0);
2751 if (atomic_dec_and_test(&rs
->rs_refcount
))
2752 lustre_free_reply_state(rs
);
2755 /* Should only be called once per req */
2756 static inline void ptlrpc_req_drop_rs(struct ptlrpc_request
*req
)
2758 if (!req
->rq_reply_state
)
2759 return; /* shouldn't occur */
2760 ptlrpc_rs_decref(req
->rq_reply_state
);
2761 req
->rq_reply_state
= NULL
;
2762 req
->rq_repmsg
= NULL
;
2765 static inline __u32
lustre_request_magic(struct ptlrpc_request
*req
)
2767 return lustre_msg_get_magic(req
->rq_reqmsg
);
2770 static inline int ptlrpc_req_get_repsize(struct ptlrpc_request
*req
)
2772 switch (req
->rq_reqmsg
->lm_magic
) {
2773 case LUSTRE_MSG_MAGIC_V2
:
2774 return req
->rq_reqmsg
->lm_repsize
;
2776 LASSERTF(0, "incorrect message magic: %08x\n",
2777 req
->rq_reqmsg
->lm_magic
);
2782 static inline int ptlrpc_send_limit_expired(struct ptlrpc_request
*req
)
2784 if (req
->rq_delay_limit
!= 0 &&
2785 time_before(cfs_time_add(req
->rq_queued_time
,
2786 cfs_time_seconds(req
->rq_delay_limit
)),
2787 cfs_time_current())) {
2793 static inline int ptlrpc_no_resend(struct ptlrpc_request
*req
)
2795 if (!req
->rq_no_resend
&& ptlrpc_send_limit_expired(req
)) {
2796 spin_lock(&req
->rq_lock
);
2797 req
->rq_no_resend
= 1;
2798 spin_unlock(&req
->rq_lock
);
2800 return req
->rq_no_resend
;
2804 ptlrpc_server_get_timeout(struct ptlrpc_service_part
*svcpt
)
2806 int at
= AT_OFF
? 0 : at_get(&svcpt
->scp_at_estimate
);
2808 return svcpt
->scp_service
->srv_watchdog_factor
*
2809 max_t(int, at
, obd_timeout
);
2812 static inline struct ptlrpc_service
*
2813 ptlrpc_req2svc(struct ptlrpc_request
*req
)
2815 return req
->rq_rqbd
->rqbd_svcpt
->scp_service
;
2818 /* ldlm/ldlm_lib.c */
2820 * Target client logic
2823 int client_obd_setup(struct obd_device
*obddev
, struct lustre_cfg
*lcfg
);
2824 int client_obd_cleanup(struct obd_device
*obddev
);
2825 int client_connect_import(const struct lu_env
*env
,
2826 struct obd_export
**exp
, struct obd_device
*obd
,
2827 struct obd_uuid
*cluuid
, struct obd_connect_data
*,
2829 int client_disconnect_export(struct obd_export
*exp
);
2830 int client_import_add_conn(struct obd_import
*imp
, struct obd_uuid
*uuid
,
2832 int client_import_del_conn(struct obd_import
*imp
, struct obd_uuid
*uuid
);
2833 int client_import_find_conn(struct obd_import
*imp
, lnet_nid_t peer
,
2834 struct obd_uuid
*uuid
);
2835 int import_set_conn_priority(struct obd_import
*imp
, struct obd_uuid
*uuid
);
2836 void client_destroy_import(struct obd_import
*imp
);
2839 /* ptlrpc/pinger.c */
2841 * Pinger API (client side only)
2844 enum timeout_event
{
2848 struct timeout_item
;
2849 typedef int (*timeout_cb_t
)(struct timeout_item
*, void *);
2850 int ptlrpc_pinger_add_import(struct obd_import
*imp
);
2851 int ptlrpc_pinger_del_import(struct obd_import
*imp
);
2852 int ptlrpc_add_timeout_client(int time
, enum timeout_event event
,
2853 timeout_cb_t cb
, void *data
,
2854 struct list_head
*obd_list
);
2855 int ptlrpc_del_timeout_client(struct list_head
*obd_list
,
2856 enum timeout_event event
);
2857 struct ptlrpc_request
*ptlrpc_prep_ping(struct obd_import
*imp
);
2858 int ptlrpc_obd_ping(struct obd_device
*obd
);
2859 void ptlrpc_pinger_ir_up(void);
2860 void ptlrpc_pinger_ir_down(void);
2862 int ptlrpc_pinger_suppress_pings(void);
2864 /* ptlrpc/ptlrpcd.c */
2865 void ptlrpcd_stop(struct ptlrpcd_ctl
*pc
, int force
);
2866 void ptlrpcd_free(struct ptlrpcd_ctl
*pc
);
2867 void ptlrpcd_wake(struct ptlrpc_request
*req
);
2868 void ptlrpcd_add_req(struct ptlrpc_request
*req
);
2869 int ptlrpcd_addref(void);
2870 void ptlrpcd_decref(void);
2872 /* ptlrpc/lproc_ptlrpc.c */
2874 * procfs output related functions
2877 const char *ll_opcode2str(__u32 opcode
);
2878 void ptlrpc_lprocfs_register_obd(struct obd_device
*obd
);
2879 void ptlrpc_lprocfs_unregister_obd(struct obd_device
*obd
);
2880 void ptlrpc_lprocfs_brw(struct ptlrpc_request
*req
, int bytes
);
2883 /* ptlrpc/llog_client.c */
2884 extern struct llog_operations llog_client_ops
;