Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
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. | |
9 | * | |
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). | |
15 | * | |
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 | |
19 | * | |
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 | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2010, 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | */ | |
36 | /** \defgroup PtlRPC Portal RPC and networking module. | |
37 | * | |
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. | |
43 | * | |
44 | * PortalRPC utilizes LNet as its transport layer. | |
45 | * | |
46 | * @{ | |
47 | */ | |
48 | ||
49 | ||
50 | #ifndef _LUSTRE_NET_H | |
51 | #define _LUSTRE_NET_H | |
52 | ||
53 | /** \defgroup net net | |
54 | * | |
55 | * @{ | |
56 | */ | |
57 | ||
1accaadf | 58 | #include "linux/lustre_net.h" |
d7e09d03 | 59 | |
9fdaf8c0 | 60 | #include "../../include/linux/libcfs/libcfs.h" |
d7e09d03 | 61 | // #include <obd.h> |
9fdaf8c0 | 62 | #include "../../include/linux/lnet/lnet.h" |
1accaadf GKH |
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" | |
d7e09d03 | 70 | |
1accaadf GKH |
71 | #include "obd_support.h" |
72 | #include "lustre_ver.h" | |
d7e09d03 PT |
73 | |
74 | /* MD flags we _always_ use */ | |
75 | #define PTLRPC_MD_OPTIONS 0 | |
76 | ||
77 | /** | |
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) | |
85 | /** | |
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 | |
91 | * RPC count. */ | |
92 | #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1)) | |
93 | ||
94 | /** | |
95 | * Define maxima for bulk I/O. | |
96 | * | |
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. | |
100 | */ | |
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) | |
104 | ||
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) | |
111 | ||
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" | |
115 | # endif | |
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" | |
118 | # endif | |
119 | # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT) | |
120 | # error "PTLRPC_MAX_BRW_SIZE too big" | |
121 | # endif | |
122 | # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT) | |
123 | # error "PTLRPC_MAX_BRW_PAGES too big" | |
124 | # endif | |
125 | ||
126 | #define PTLRPC_NTHRS_INIT 2 | |
127 | ||
128 | /** | |
129 | * Buffer Constants | |
130 | * | |
131 | * Constants determine how memory is used to buffer incoming service requests. | |
132 | * | |
133 | * ?_NBUFS # buffers to allocate when growing the pool | |
134 | * ?_BUFSIZE # bytes in a single request buffer | |
135 | * ?_MAXREQSIZE # maximum request service will receive | |
136 | * | |
137 | * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk | |
138 | * of ?_NBUFS is added to the pool. | |
139 | * | |
140 | * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are | |
141 | * considered full when less than ?_MAXREQSIZE is left in them. | |
142 | */ | |
143 | /** | |
144 | * Thread Constants | |
145 | * | |
146 | * Constants determine how threads are created for ptlrpc service. | |
147 | * | |
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. | |
167 | * | |
168 | * Examples | |
169 | * | |
170 | * #define MDS_NTHRS_INIT 2 | |
171 | * #define MDS_NTHRS_BASE 64 | |
172 | * #define MDS_NTHRS_FACTOR 8 | |
173 | * #define MDS_NTHRS_MAX 1024 | |
174 | * | |
175 | * Example 1): | |
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 | |
179 | * partition is: | |
180 | * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96 | |
181 | * | |
182 | * Total number of threads for the service is: | |
183 | * 96 * partitions(4) = 384 | |
184 | * | |
185 | * Example 2): | |
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 | |
189 | * partition is: | |
190 | * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128 | |
191 | * | |
192 | * Total number of threads for the service is: | |
193 | * 128 * partitions(4) = 512 | |
194 | * | |
195 | * Example 3): | |
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 | |
199 | * partition is: | |
200 | * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160 | |
201 | * | |
202 | * Total number of threads for the service is: | |
203 | * 160 * partitions(8) = 1280 | |
204 | * | |
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 | |
208 | * | |
209 | * Example 4): | |
210 | * --------------------------------------------------------------------- | |
211 | * Server(C) have a thousand of cores and user configured it to 32 partitions | |
212 | * MDS_NTHRS_BASE(64) * 32 = 2048 | |
213 | * | |
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. | |
217 | * | |
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. | |
221 | * | |
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 | |
225 | * top of this subset | |
226 | * b) bind service threads on a few partitions, see modparameters of | |
227 | * MDS and OSS for details | |
228 | * | |
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. | |
232 | * | |
233 | */ | |
234 | ||
235 | /* | |
236 | * LDLM threads constants: | |
237 | * | |
238 | * Given 8 as factor and 24 as base threads number | |
239 | * | |
240 | * example 1) | |
241 | * On 4-core machine we will have 24 + 8 * 4 = 56 threads. | |
242 | * | |
243 | * example 2) | |
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. | |
246 | * | |
247 | * example 3) | |
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. | |
251 | * | |
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 | |
254 | */ | |
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) | |
259 | ||
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) | |
266 | ||
e55c4476 JN |
267 | #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */ |
268 | ||
d7e09d03 | 269 | #define OST_MAXREQSIZE (5 * 1024) |
d7e09d03 PT |
270 | |
271 | /* Macro to hide a typecast. */ | |
272 | #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args) | |
273 | ||
274 | /** | |
275 | * Structure to single define portal connection. | |
276 | */ | |
277 | struct ptlrpc_connection { | |
278 | /** linkage for connections hash table */ | |
279 | struct hlist_node c_hash; | |
280 | /** Our own lnet nid for this connection */ | |
281 | lnet_nid_t c_self; | |
282 | /** Remote side nid for this connection */ | |
283 | lnet_process_id_t c_peer; | |
284 | /** UUID of the other side */ | |
285 | struct obd_uuid c_remote_uuid; | |
286 | /** reference counter for this connection */ | |
287 | atomic_t c_refcount; | |
288 | }; | |
289 | ||
290 | /** Client definition for PortalRPC */ | |
291 | struct ptlrpc_client { | |
292 | /** What lnet portal does this client send messages to by default */ | |
293 | __u32 cli_request_portal; | |
294 | /** What portal do we expect replies on */ | |
295 | __u32 cli_reply_portal; | |
296 | /** Name of the client */ | |
297 | char *cli_name; | |
298 | }; | |
299 | ||
300 | /** state flags of requests */ | |
301 | /* XXX only ones left are those used by the bulk descs as well! */ | |
302 | #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */ | |
303 | #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */ | |
304 | ||
305 | #define REQ_MAX_ACK_LOCKS 8 | |
306 | ||
307 | union ptlrpc_async_args { | |
308 | /** | |
309 | * Scratchpad for passing args to completion interpreter. Users | |
310 | * cast to the struct of their choosing, and CLASSERT that this is | |
311 | * big enough. For _tons_ of context, OBD_ALLOC a struct and store | |
312 | * a pointer to it here. The pointer_arg ensures this struct is at | |
313 | * least big enough for that. | |
314 | */ | |
315 | void *pointer_arg[11]; | |
316 | __u64 space[7]; | |
317 | }; | |
318 | ||
319 | struct ptlrpc_request_set; | |
320 | typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int); | |
321 | typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *); | |
322 | ||
323 | /** | |
324 | * Definition of request set structure. | |
325 | * Request set is a list of requests (not necessary to the same target) that | |
326 | * once populated with RPCs could be sent in parallel. | |
327 | * There are two kinds of request sets. General purpose and with dedicated | |
328 | * serving thread. Example of the latter is ptlrpcd set. | |
329 | * For general purpose sets once request set started sending it is impossible | |
330 | * to add new requests to such set. | |
331 | * Provides a way to call "completion callbacks" when all requests in the set | |
332 | * returned. | |
333 | */ | |
334 | struct ptlrpc_request_set { | |
335 | atomic_t set_refcount; | |
336 | /** number of in queue requests */ | |
337 | atomic_t set_new_count; | |
338 | /** number of uncompleted requests */ | |
339 | atomic_t set_remaining; | |
340 | /** wait queue to wait on for request events */ | |
341 | wait_queue_head_t set_waitq; | |
342 | wait_queue_head_t *set_wakeup_ptr; | |
343 | /** List of requests in the set */ | |
344 | struct list_head set_requests; | |
345 | /** | |
346 | * List of completion callbacks to be called when the set is completed | |
347 | * This is only used if \a set_interpret is NULL. | |
348 | * Links struct ptlrpc_set_cbdata. | |
349 | */ | |
350 | struct list_head set_cblist; | |
351 | /** Completion callback, if only one. */ | |
352 | set_interpreter_func set_interpret; | |
353 | /** opaq argument passed to completion \a set_interpret callback. */ | |
354 | void *set_arg; | |
355 | /** | |
356 | * Lock for \a set_new_requests manipulations | |
357 | * locked so that any old caller can communicate requests to | |
358 | * the set holder who can then fold them into the lock-free set | |
359 | */ | |
360 | spinlock_t set_new_req_lock; | |
361 | /** List of new yet unsent requests. Only used with ptlrpcd now. */ | |
362 | struct list_head set_new_requests; | |
363 | ||
364 | /** rq_status of requests that have been freed already */ | |
365 | int set_rc; | |
366 | /** Additional fields used by the flow control extension */ | |
367 | /** Maximum number of RPCs in flight */ | |
368 | int set_max_inflight; | |
369 | /** Callback function used to generate RPCs */ | |
370 | set_producer_func set_producer; | |
371 | /** opaq argument passed to the producer callback */ | |
372 | void *set_producer_arg; | |
373 | }; | |
374 | ||
375 | /** | |
376 | * Description of a single ptrlrpc_set callback | |
377 | */ | |
378 | struct ptlrpc_set_cbdata { | |
379 | /** List linkage item */ | |
380 | struct list_head psc_item; | |
381 | /** Pointer to interpreting function */ | |
382 | set_interpreter_func psc_interpret; | |
383 | /** Opaq argument to pass to the callback */ | |
384 | void *psc_data; | |
385 | }; | |
386 | ||
387 | struct ptlrpc_bulk_desc; | |
388 | struct ptlrpc_service_part; | |
389 | struct ptlrpc_service; | |
390 | ||
391 | /** | |
392 | * ptlrpc callback & work item stuff | |
393 | */ | |
394 | struct ptlrpc_cb_id { | |
395 | void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */ | |
396 | void *cbid_arg; /* additional arg */ | |
397 | }; | |
398 | ||
399 | /** Maximum number of locks to fit into reply state */ | |
400 | #define RS_MAX_LOCKS 8 | |
401 | #define RS_DEBUG 0 | |
402 | ||
403 | /** | |
404 | * Structure to define reply state on the server | |
405 | * Reply state holds various reply message information. Also for "difficult" | |
406 | * replies (rep-ack case) we store the state after sending reply and wait | |
407 | * for the client to acknowledge the reception. In these cases locks could be | |
408 | * added to the state for replay/failover consistency guarantees. | |
409 | */ | |
410 | struct ptlrpc_reply_state { | |
411 | /** Callback description */ | |
412 | struct ptlrpc_cb_id rs_cb_id; | |
413 | /** Linkage for list of all reply states in a system */ | |
414 | struct list_head rs_list; | |
415 | /** Linkage for list of all reply states on same export */ | |
416 | struct list_head rs_exp_list; | |
417 | /** Linkage for list of all reply states for same obd */ | |
418 | struct list_head rs_obd_list; | |
419 | #if RS_DEBUG | |
420 | struct list_head rs_debug_list; | |
421 | #endif | |
422 | /** A spinlock to protect the reply state flags */ | |
423 | spinlock_t rs_lock; | |
424 | /** Reply state flags */ | |
425 | unsigned long rs_difficult:1; /* ACK/commit stuff */ | |
426 | unsigned long rs_no_ack:1; /* no ACK, even for | |
427 | difficult requests */ | |
428 | unsigned long rs_scheduled:1; /* being handled? */ | |
429 | unsigned long rs_scheduled_ever:1;/* any schedule attempts? */ | |
430 | unsigned long rs_handled:1; /* been handled yet? */ | |
431 | unsigned long rs_on_net:1; /* reply_out_callback pending? */ | |
432 | unsigned long rs_prealloc:1; /* rs from prealloc list */ | |
433 | unsigned long rs_committed:1;/* the transaction was committed | |
434 | and the rs was dispatched | |
435 | by ptlrpc_commit_replies */ | |
436 | /** Size of the state */ | |
437 | int rs_size; | |
438 | /** opcode */ | |
439 | __u32 rs_opc; | |
440 | /** Transaction number */ | |
441 | __u64 rs_transno; | |
442 | /** xid */ | |
443 | __u64 rs_xid; | |
444 | struct obd_export *rs_export; | |
445 | struct ptlrpc_service_part *rs_svcpt; | |
446 | /** Lnet metadata handle for the reply */ | |
447 | lnet_handle_md_t rs_md_h; | |
448 | atomic_t rs_refcount; | |
449 | ||
17891183 | 450 | /** Context for the service thread */ |
d7e09d03 PT |
451 | struct ptlrpc_svc_ctx *rs_svc_ctx; |
452 | /** Reply buffer (actually sent to the client), encoded if needed */ | |
453 | struct lustre_msg *rs_repbuf; /* wrapper */ | |
454 | /** Size of the reply buffer */ | |
455 | int rs_repbuf_len; /* wrapper buf length */ | |
456 | /** Size of the reply message */ | |
457 | int rs_repdata_len; /* wrapper msg length */ | |
458 | /** | |
bd9070cb | 459 | * Actual reply message. Its content is encrypted (if needed) to |
d7e09d03 | 460 | * produce reply buffer for actual sending. In simple case |
bd9070cb | 461 | * of no network encryption we just set \a rs_repbuf to \a rs_msg |
d7e09d03 PT |
462 | */ |
463 | struct lustre_msg *rs_msg; /* reply message */ | |
464 | ||
465 | /** Number of locks awaiting client ACK */ | |
466 | int rs_nlocks; | |
467 | /** Handles of locks awaiting client reply ACK */ | |
468 | struct lustre_handle rs_locks[RS_MAX_LOCKS]; | |
469 | /** Lock modes of locks in \a rs_locks */ | |
470 | ldlm_mode_t rs_modes[RS_MAX_LOCKS]; | |
471 | }; | |
472 | ||
473 | struct ptlrpc_thread; | |
474 | ||
475 | /** RPC stages */ | |
476 | enum rq_phase { | |
477 | RQ_PHASE_NEW = 0xebc0de00, | |
478 | RQ_PHASE_RPC = 0xebc0de01, | |
479 | RQ_PHASE_BULK = 0xebc0de02, | |
480 | RQ_PHASE_INTERPRET = 0xebc0de03, | |
481 | RQ_PHASE_COMPLETE = 0xebc0de04, | |
482 | RQ_PHASE_UNREGISTERING = 0xebc0de05, | |
483 | RQ_PHASE_UNDEFINED = 0xebc0de06 | |
484 | }; | |
485 | ||
486 | /** Type of request interpreter call-back */ | |
487 | typedef int (*ptlrpc_interpterer_t)(const struct lu_env *env, | |
488 | struct ptlrpc_request *req, | |
489 | void *arg, int rc); | |
490 | ||
491 | /** | |
492 | * Definition of request pool structure. | |
493 | * The pool is used to store empty preallocated requests for the case | |
494 | * when we would actually need to send something without performing | |
495 | * any allocations (to avoid e.g. OOM). | |
496 | */ | |
497 | struct ptlrpc_request_pool { | |
498 | /** Locks the list */ | |
499 | spinlock_t prp_lock; | |
500 | /** list of ptlrpc_request structs */ | |
501 | struct list_head prp_req_list; | |
17891183 | 502 | /** Maximum message size that would fit into a request from this pool */ |
d7e09d03 PT |
503 | int prp_rq_size; |
504 | /** Function to allocate more requests for this pool */ | |
505 | void (*prp_populate)(struct ptlrpc_request_pool *, int); | |
506 | }; | |
507 | ||
508 | struct lu_context; | |
509 | struct lu_env; | |
510 | ||
511 | struct ldlm_lock; | |
512 | ||
513 | /** | |
514 | * \defgroup nrs Network Request Scheduler | |
515 | * @{ | |
516 | */ | |
517 | struct ptlrpc_nrs_policy; | |
518 | struct ptlrpc_nrs_resource; | |
519 | struct ptlrpc_nrs_request; | |
520 | ||
521 | /** | |
522 | * NRS control operations. | |
523 | * | |
524 | * These are common for all policies. | |
525 | */ | |
526 | enum ptlrpc_nrs_ctl { | |
527 | /** | |
528 | * Not a valid opcode. | |
529 | */ | |
530 | PTLRPC_NRS_CTL_INVALID, | |
531 | /** | |
532 | * Activate the policy. | |
533 | */ | |
534 | PTLRPC_NRS_CTL_START, | |
535 | /** | |
536 | * Reserved for multiple primary policies, which may be a possibility | |
537 | * in the future. | |
538 | */ | |
539 | PTLRPC_NRS_CTL_STOP, | |
540 | /** | |
541 | * Policies can start using opcodes from this value and onwards for | |
542 | * their own purposes; the assigned value itself is arbitrary. | |
543 | */ | |
544 | PTLRPC_NRS_CTL_1ST_POL_SPEC = 0x20, | |
545 | }; | |
546 | ||
547 | /** | |
548 | * ORR policy operations | |
549 | */ | |
550 | enum nrs_ctl_orr { | |
551 | NRS_CTL_ORR_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC, | |
552 | NRS_CTL_ORR_WR_QUANTUM, | |
553 | NRS_CTL_ORR_RD_OFF_TYPE, | |
554 | NRS_CTL_ORR_WR_OFF_TYPE, | |
555 | NRS_CTL_ORR_RD_SUPP_REQ, | |
556 | NRS_CTL_ORR_WR_SUPP_REQ, | |
557 | }; | |
558 | ||
559 | /** | |
560 | * NRS policy operations. | |
561 | * | |
562 | * These determine the behaviour of a policy, and are called in response to | |
563 | * NRS core events. | |
564 | */ | |
565 | struct ptlrpc_nrs_pol_ops { | |
566 | /** | |
567 | * Called during policy registration; this operation is optional. | |
568 | * | |
569 | * \param[in,out] policy The policy being initialized | |
570 | */ | |
571 | int (*op_policy_init) (struct ptlrpc_nrs_policy *policy); | |
572 | /** | |
573 | * Called during policy unregistration; this operation is optional. | |
574 | * | |
575 | * \param[in,out] policy The policy being unregistered/finalized | |
576 | */ | |
577 | void (*op_policy_fini) (struct ptlrpc_nrs_policy *policy); | |
578 | /** | |
579 | * Called when activating a policy via lprocfs; policies allocate and | |
580 | * initialize their resources here; this operation is optional. | |
581 | * | |
582 | * \param[in,out] policy The policy being started | |
583 | * | |
584 | * \see nrs_policy_start_locked() | |
585 | */ | |
586 | int (*op_policy_start) (struct ptlrpc_nrs_policy *policy); | |
587 | /** | |
588 | * Called when deactivating a policy via lprocfs; policies deallocate | |
589 | * their resources here; this operation is optional | |
590 | * | |
591 | * \param[in,out] policy The policy being stopped | |
592 | * | |
593 | * \see nrs_policy_stop0() | |
594 | */ | |
595 | void (*op_policy_stop) (struct ptlrpc_nrs_policy *policy); | |
596 | /** | |
597 | * Used for policy-specific operations; i.e. not generic ones like | |
598 | * \e PTLRPC_NRS_CTL_START and \e PTLRPC_NRS_CTL_GET_INFO; analogous | |
599 | * to an ioctl; this operation is optional. | |
600 | * | |
601 | * \param[in,out] policy The policy carrying out operation \a opc | |
602 | * \param[in] opc The command operation being carried out | |
603 | * \param[in,out] arg An generic buffer for communication between the | |
604 | * user and the control operation | |
605 | * | |
606 | * \retval -ve error | |
607 | * \retval 0 success | |
608 | * | |
609 | * \see ptlrpc_nrs_policy_control() | |
610 | */ | |
611 | int (*op_policy_ctl) (struct ptlrpc_nrs_policy *policy, | |
612 | enum ptlrpc_nrs_ctl opc, void *arg); | |
613 | ||
614 | /** | |
615 | * Called when obtaining references to the resources of the resource | |
616 | * hierarchy for a request that has arrived for handling at the PTLRPC | |
617 | * service. Policies should return -ve for requests they do not wish | |
618 | * to handle. This operation is mandatory. | |
619 | * | |
620 | * \param[in,out] policy The policy we're getting resources for. | |
621 | * \param[in,out] nrq The request we are getting resources for. | |
622 | * \param[in] parent The parent resource of the resource being | |
623 | * requested; set to NULL if none. | |
624 | * \param[out] resp The resource is to be returned here; the | |
625 | * fallback policy in an NRS head should | |
626 | * \e always return a non-NULL pointer value. | |
627 | * \param[in] moving_req When set, signifies that this is an attempt | |
628 | * to obtain resources for a request being moved | |
629 | * to the high-priority NRS head by | |
630 | * ldlm_lock_reorder_req(). | |
631 | * This implies two things: | |
632 | * 1. We are under obd_export::exp_rpc_lock and | |
633 | * so should not sleep. | |
634 | * 2. We should not perform non-idempotent or can | |
635 | * skip performing idempotent operations that | |
636 | * were carried out when resources were first | |
637 | * taken for the request when it was initialized | |
638 | * in ptlrpc_nrs_req_initialize(). | |
639 | * | |
640 | * \retval 0, +ve The level of the returned resource in the resource | |
641 | * hierarchy; currently only 0 (for a non-leaf resource) | |
642 | * and 1 (for a leaf resource) are supported by the | |
643 | * framework. | |
644 | * \retval -ve error | |
645 | * | |
646 | * \see ptlrpc_nrs_req_initialize() | |
647 | * \see ptlrpc_nrs_hpreq_add_nolock() | |
648 | * \see ptlrpc_nrs_req_hp_move() | |
649 | */ | |
650 | int (*op_res_get) (struct ptlrpc_nrs_policy *policy, | |
651 | struct ptlrpc_nrs_request *nrq, | |
652 | const struct ptlrpc_nrs_resource *parent, | |
653 | struct ptlrpc_nrs_resource **resp, | |
654 | bool moving_req); | |
655 | /** | |
656 | * Called when releasing references taken for resources in the resource | |
657 | * hierarchy for the request; this operation is optional. | |
658 | * | |
659 | * \param[in,out] policy The policy the resource belongs to | |
660 | * \param[in] res The resource to be freed | |
661 | * | |
662 | * \see ptlrpc_nrs_req_finalize() | |
663 | * \see ptlrpc_nrs_hpreq_add_nolock() | |
664 | * \see ptlrpc_nrs_req_hp_move() | |
665 | */ | |
666 | void (*op_res_put) (struct ptlrpc_nrs_policy *policy, | |
667 | const struct ptlrpc_nrs_resource *res); | |
668 | ||
669 | /** | |
670 | * Obtains a request for handling from the policy, and optionally | |
671 | * removes the request from the policy; this operation is mandatory. | |
672 | * | |
673 | * \param[in,out] policy The policy to poll | |
674 | * \param[in] peek When set, signifies that we just want to | |
675 | * examine the request, and not handle it, so the | |
676 | * request is not removed from the policy. | |
677 | * \param[in] force When set, it will force a policy to return a | |
678 | * request if it has one queued. | |
679 | * | |
680 | * \retval NULL No request available for handling | |
681 | * \retval valid-pointer The request polled for handling | |
682 | * | |
683 | * \see ptlrpc_nrs_req_get_nolock() | |
684 | */ | |
685 | struct ptlrpc_nrs_request * | |
686 | (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek, | |
687 | bool force); | |
688 | /** | |
689 | * Called when attempting to add a request to a policy for later | |
690 | * handling; this operation is mandatory. | |
691 | * | |
692 | * \param[in,out] policy The policy on which to enqueue \a nrq | |
693 | * \param[in,out] nrq The request to enqueue | |
694 | * | |
695 | * \retval 0 success | |
696 | * \retval != 0 error | |
697 | * | |
698 | * \see ptlrpc_nrs_req_add_nolock() | |
699 | */ | |
700 | int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy, | |
701 | struct ptlrpc_nrs_request *nrq); | |
702 | /** | |
703 | * Removes a request from the policy's set of pending requests. Normally | |
704 | * called after a request has been polled successfully from the policy | |
705 | * for handling; this operation is mandatory. | |
706 | * | |
707 | * \param[in,out] policy The policy the request \a nrq belongs to | |
708 | * \param[in,out] nrq The request to dequeue | |
709 | * | |
710 | * \see ptlrpc_nrs_req_del_nolock() | |
711 | */ | |
712 | void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy, | |
713 | struct ptlrpc_nrs_request *nrq); | |
714 | /** | |
715 | * Called after the request being carried out. Could be used for | |
716 | * job/resource control; this operation is optional. | |
717 | * | |
718 | * \param[in,out] policy The policy which is stopping to handle request | |
719 | * \a nrq | |
720 | * \param[in,out] nrq The request | |
721 | * | |
5e42bc9d | 722 | * \pre assert_spin_locked(&svcpt->scp_req_lock) |
d7e09d03 PT |
723 | * |
724 | * \see ptlrpc_nrs_req_stop_nolock() | |
725 | */ | |
726 | void (*op_req_stop) (struct ptlrpc_nrs_policy *policy, | |
727 | struct ptlrpc_nrs_request *nrq); | |
728 | /** | |
729 | * Registers the policy's lprocfs interface with a PTLRPC service. | |
730 | * | |
731 | * \param[in] svc The service | |
732 | * | |
733 | * \retval 0 success | |
734 | * \retval != 0 error | |
735 | */ | |
736 | int (*op_lprocfs_init) (struct ptlrpc_service *svc); | |
737 | /** | |
738 | * Unegisters the policy's lprocfs interface with a PTLRPC service. | |
739 | * | |
740 | * In cases of failed policy registration in | |
741 | * \e ptlrpc_nrs_policy_register(), this function may be called for a | |
742 | * service which has not registered the policy successfully, so | |
743 | * implementations of this method should make sure their operations are | |
744 | * safe in such cases. | |
745 | * | |
746 | * \param[in] svc The service | |
747 | */ | |
748 | void (*op_lprocfs_fini) (struct ptlrpc_service *svc); | |
749 | }; | |
750 | ||
751 | /** | |
752 | * Policy flags | |
753 | */ | |
754 | enum nrs_policy_flags { | |
755 | /** | |
756 | * Fallback policy, use this flag only on a single supported policy per | |
757 | * service. The flag cannot be used on policies that use | |
758 | * \e PTLRPC_NRS_FL_REG_EXTERN | |
759 | */ | |
760 | PTLRPC_NRS_FL_FALLBACK = (1 << 0), | |
761 | /** | |
762 | * Start policy immediately after registering. | |
763 | */ | |
764 | PTLRPC_NRS_FL_REG_START = (1 << 1), | |
765 | /** | |
766 | * This is a policy registering from a module different to the one NRS | |
767 | * core ships in (currently ptlrpc). | |
768 | */ | |
769 | PTLRPC_NRS_FL_REG_EXTERN = (1 << 2), | |
770 | }; | |
771 | ||
772 | /** | |
773 | * NRS queue type. | |
774 | * | |
775 | * Denotes whether an NRS instance is for handling normal or high-priority | |
776 | * RPCs, or whether an operation pertains to one or both of the NRS instances | |
777 | * in a service. | |
778 | */ | |
779 | enum ptlrpc_nrs_queue_type { | |
780 | PTLRPC_NRS_QUEUE_REG = (1 << 0), | |
781 | PTLRPC_NRS_QUEUE_HP = (1 << 1), | |
782 | PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP) | |
783 | }; | |
784 | ||
785 | /** | |
786 | * NRS head | |
787 | * | |
788 | * A PTLRPC service has at least one NRS head instance for handling normal | |
789 | * priority RPCs, and may optionally have a second NRS head instance for | |
790 | * handling high-priority RPCs. Each NRS head maintains a list of available | |
791 | * policies, of which one and only one policy is acting as the fallback policy, | |
792 | * and optionally a different policy may be acting as the primary policy. For | |
793 | * all RPCs handled by this NRS head instance, NRS core will first attempt to | |
794 | * enqueue the RPC using the primary policy (if any). The fallback policy is | |
795 | * used in the following cases: | |
796 | * - when there was no primary policy in the | |
797 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request | |
798 | * was initialized. | |
799 | * - when the primary policy that was at the | |
800 | * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the | |
801 | * RPC was initialized, denoted it did not wish, or for some other reason was | |
802 | * not able to handle the request, by returning a non-valid NRS resource | |
803 | * reference. | |
804 | * - when the primary policy that was at the | |
805 | * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the | |
806 | * RPC was initialized, fails later during the request enqueueing stage. | |
807 | * | |
808 | * \see nrs_resource_get_safe() | |
809 | * \see nrs_request_enqueue() | |
810 | */ | |
811 | struct ptlrpc_nrs { | |
812 | spinlock_t nrs_lock; | |
813 | /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */ | |
814 | /** | |
815 | * List of registered policies | |
816 | */ | |
817 | struct list_head nrs_policy_list; | |
818 | /** | |
819 | * List of policies with queued requests. Policies that have any | |
820 | * outstanding requests are queued here, and this list is queried | |
821 | * in a round-robin manner from NRS core when obtaining a request | |
822 | * for handling. This ensures that requests from policies that at some | |
823 | * point transition away from the | |
824 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained. | |
825 | */ | |
826 | struct list_head nrs_policy_queued; | |
827 | /** | |
828 | * Service partition for this NRS head | |
829 | */ | |
830 | struct ptlrpc_service_part *nrs_svcpt; | |
831 | /** | |
832 | * Primary policy, which is the preferred policy for handling RPCs | |
833 | */ | |
834 | struct ptlrpc_nrs_policy *nrs_policy_primary; | |
835 | /** | |
836 | * Fallback policy, which is the backup policy for handling RPCs | |
837 | */ | |
838 | struct ptlrpc_nrs_policy *nrs_policy_fallback; | |
839 | /** | |
840 | * This NRS head handles either HP or regular requests | |
841 | */ | |
842 | enum ptlrpc_nrs_queue_type nrs_queue_type; | |
843 | /** | |
844 | * # queued requests from all policies in this NRS head | |
845 | */ | |
846 | unsigned long nrs_req_queued; | |
847 | /** | |
848 | * # scheduled requests from all policies in this NRS head | |
849 | */ | |
850 | unsigned long nrs_req_started; | |
851 | /** | |
852 | * # policies on this NRS | |
853 | */ | |
854 | unsigned nrs_num_pols; | |
855 | /** | |
856 | * This NRS head is in progress of starting a policy | |
857 | */ | |
858 | unsigned nrs_policy_starting:1; | |
859 | /** | |
860 | * In progress of shutting down the whole NRS head; used during | |
861 | * unregistration | |
862 | */ | |
863 | unsigned nrs_stopping:1; | |
864 | }; | |
865 | ||
866 | #define NRS_POL_NAME_MAX 16 | |
867 | ||
868 | struct ptlrpc_nrs_pol_desc; | |
869 | ||
870 | /** | |
871 | * Service compatibility predicate; this determines whether a policy is adequate | |
872 | * for handling RPCs of a particular PTLRPC service. | |
873 | * | |
874 | * XXX:This should give the same result during policy registration and | |
875 | * unregistration, and for all partitions of a service; so the result should not | |
876 | * depend on temporal service or other properties, that may influence the | |
877 | * result. | |
878 | */ | |
879 | typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc, | |
880 | const struct ptlrpc_nrs_pol_desc *desc); | |
881 | ||
882 | struct ptlrpc_nrs_pol_conf { | |
883 | /** | |
884 | * Human-readable policy name | |
885 | */ | |
886 | char nc_name[NRS_POL_NAME_MAX]; | |
887 | /** | |
888 | * NRS operations for this policy | |
889 | */ | |
890 | const struct ptlrpc_nrs_pol_ops *nc_ops; | |
891 | /** | |
892 | * Service compatibility predicate | |
893 | */ | |
894 | nrs_pol_desc_compat_t nc_compat; | |
895 | /** | |
896 | * Set for policies that support a single ptlrpc service, i.e. ones that | |
897 | * have \a pd_compat set to nrs_policy_compat_one(). The variable value | |
898 | * depicts the name of the single service that such policies are | |
899 | * compatible with. | |
900 | */ | |
901 | const char *nc_compat_svc_name; | |
902 | /** | |
903 | * Owner module for this policy descriptor; policies registering from a | |
904 | * different module to the one the NRS framework is held within | |
905 | * (currently ptlrpc), should set this field to THIS_MODULE. | |
906 | */ | |
c34d9cd8 | 907 | struct module *nc_owner; |
d7e09d03 | 908 | /** |
bd9070cb | 909 | * Policy registration flags; a bitmask of \e nrs_policy_flags |
d7e09d03 PT |
910 | */ |
911 | unsigned nc_flags; | |
912 | }; | |
913 | ||
914 | /** | |
915 | * NRS policy registering descriptor | |
916 | * | |
917 | * Is used to hold a description of a policy that can be passed to NRS core in | |
918 | * order to register the policy with NRS heads in different PTLRPC services. | |
919 | */ | |
920 | struct ptlrpc_nrs_pol_desc { | |
921 | /** | |
922 | * Human-readable policy name | |
923 | */ | |
924 | char pd_name[NRS_POL_NAME_MAX]; | |
925 | /** | |
926 | * Link into nrs_core::nrs_policies | |
927 | */ | |
928 | struct list_head pd_list; | |
929 | /** | |
930 | * NRS operations for this policy | |
931 | */ | |
932 | const struct ptlrpc_nrs_pol_ops *pd_ops; | |
933 | /** | |
934 | * Service compatibility predicate | |
935 | */ | |
936 | nrs_pol_desc_compat_t pd_compat; | |
937 | /** | |
938 | * Set for policies that are compatible with only one PTLRPC service. | |
939 | * | |
940 | * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name | |
941 | */ | |
942 | const char *pd_compat_svc_name; | |
943 | /** | |
944 | * Owner module for this policy descriptor. | |
945 | * | |
946 | * We need to hold a reference to the module whenever we might make use | |
947 | * of any of the module's contents, i.e. | |
948 | * - If one or more instances of the policy are at a state where they | |
949 | * might be handling a request, i.e. | |
950 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or | |
951 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to | |
952 | * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference | |
953 | * is taken on the module when | |
954 | * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it | |
955 | * becomes 0, so that we hold only one reference to the module maximum | |
956 | * at any time. | |
957 | * | |
958 | * We do not need to hold a reference to the module, even though we | |
959 | * might use code and data from the module, in the following cases: | |
960 | * - During external policy registration, because this should happen in | |
961 | * the module's init() function, in which case the module is safe from | |
962 | * removal because a reference is being held on the module by the | |
963 | * kernel, and iirc kmod (and I guess module-init-tools also) will | |
964 | * serialize any racing processes properly anyway. | |
965 | * - During external policy unregistration, because this should happen | |
966 | * in a module's exit() function, and any attempts to start a policy | |
967 | * instance would need to take a reference on the module, and this is | |
968 | * not possible once we have reached the point where the exit() | |
969 | * handler is called. | |
970 | * - During service registration and unregistration, as service setup | |
971 | * and cleanup, and policy registration, unregistration and policy | |
972 | * instance starting, are serialized by \e nrs_core::nrs_mutex, so | |
973 | * as long as users adhere to the convention of registering policies | |
974 | * in init() and unregistering them in module exit() functions, there | |
975 | * should not be a race between these operations. | |
976 | * - During any policy-specific lprocfs operations, because a reference | |
977 | * is held by the kernel on a proc entry that has been entered by a | |
978 | * syscall, so as long as proc entries are removed during unregistration time, | |
979 | * then unregistration and lprocfs operations will be properly | |
980 | * serialized. | |
981 | */ | |
c34d9cd8 | 982 | struct module *pd_owner; |
d7e09d03 PT |
983 | /** |
984 | * Bitmask of \e nrs_policy_flags | |
985 | */ | |
986 | unsigned pd_flags; | |
987 | /** | |
988 | * # of references on this descriptor | |
989 | */ | |
990 | atomic_t pd_refs; | |
991 | }; | |
992 | ||
993 | /** | |
994 | * NRS policy state | |
995 | * | |
996 | * Policies transition from one state to the other during their lifetime | |
997 | */ | |
998 | enum ptlrpc_nrs_pol_state { | |
999 | /** | |
1000 | * Not a valid policy state. | |
1001 | */ | |
1002 | NRS_POL_STATE_INVALID, | |
1003 | /** | |
1004 | * Policies are at this state either at the start of their life, or | |
1005 | * transition here when the user selects a different policy to act | |
1006 | * as the primary one. | |
1007 | */ | |
1008 | NRS_POL_STATE_STOPPED, | |
1009 | /** | |
1010 | * Policy is progress of stopping | |
1011 | */ | |
1012 | NRS_POL_STATE_STOPPING, | |
1013 | /** | |
1014 | * Policy is in progress of starting | |
1015 | */ | |
1016 | NRS_POL_STATE_STARTING, | |
1017 | /** | |
1018 | * A policy is in this state in two cases: | |
1019 | * - it is the fallback policy, which is always in this state. | |
1020 | * - it has been activated by the user; i.e. it is the primary policy, | |
1021 | */ | |
1022 | NRS_POL_STATE_STARTED, | |
1023 | }; | |
1024 | ||
1025 | /** | |
1026 | * NRS policy information | |
1027 | * | |
1028 | * Used for obtaining information for the status of a policy via lprocfs | |
1029 | */ | |
1030 | struct ptlrpc_nrs_pol_info { | |
1031 | /** | |
1032 | * Policy name | |
1033 | */ | |
1034 | char pi_name[NRS_POL_NAME_MAX]; | |
1035 | /** | |
1036 | * Current policy state | |
1037 | */ | |
1038 | enum ptlrpc_nrs_pol_state pi_state; | |
1039 | /** | |
1040 | * # RPCs enqueued for later dispatching by the policy | |
1041 | */ | |
1042 | long pi_req_queued; | |
1043 | /** | |
1044 | * # RPCs started for dispatch by the policy | |
1045 | */ | |
1046 | long pi_req_started; | |
1047 | /** | |
1048 | * Is this a fallback policy? | |
1049 | */ | |
1050 | unsigned pi_fallback:1; | |
1051 | }; | |
1052 | ||
1053 | /** | |
1054 | * NRS policy | |
1055 | * | |
1056 | * There is one instance of this for each policy in each NRS head of each | |
1057 | * PTLRPC service partition. | |
1058 | */ | |
1059 | struct ptlrpc_nrs_policy { | |
1060 | /** | |
1061 | * Linkage into the NRS head's list of policies, | |
1062 | * ptlrpc_nrs:nrs_policy_list | |
1063 | */ | |
1064 | struct list_head pol_list; | |
1065 | /** | |
1066 | * Linkage into the NRS head's list of policies with enqueued | |
1067 | * requests ptlrpc_nrs:nrs_policy_queued | |
1068 | */ | |
1069 | struct list_head pol_list_queued; | |
1070 | /** | |
1071 | * Current state of this policy | |
1072 | */ | |
1073 | enum ptlrpc_nrs_pol_state pol_state; | |
1074 | /** | |
1075 | * Bitmask of nrs_policy_flags | |
1076 | */ | |
1077 | unsigned pol_flags; | |
1078 | /** | |
1079 | * # RPCs enqueued for later dispatching by the policy | |
1080 | */ | |
1081 | long pol_req_queued; | |
1082 | /** | |
1083 | * # RPCs started for dispatch by the policy | |
1084 | */ | |
1085 | long pol_req_started; | |
1086 | /** | |
1087 | * Usage Reference count taken on the policy instance | |
1088 | */ | |
1089 | long pol_ref; | |
1090 | /** | |
1091 | * The NRS head this policy has been created at | |
1092 | */ | |
1093 | struct ptlrpc_nrs *pol_nrs; | |
1094 | /** | |
1095 | * Private policy data; varies by policy type | |
1096 | */ | |
1097 | void *pol_private; | |
1098 | /** | |
1099 | * Policy descriptor for this policy instance. | |
1100 | */ | |
1101 | struct ptlrpc_nrs_pol_desc *pol_desc; | |
1102 | }; | |
1103 | ||
1104 | /** | |
1105 | * NRS resource | |
1106 | * | |
1107 | * Resources are embedded into two types of NRS entities: | |
1108 | * - Inside NRS policies, in the policy's private data in | |
1109 | * ptlrpc_nrs_policy::pol_private | |
1110 | * - In objects that act as prime-level scheduling entities in different NRS | |
1111 | * policies; e.g. on a policy that performs round robin or similar order | |
1112 | * scheduling across client NIDs, there would be one NRS resource per unique | |
1113 | * client NID. On a policy which performs round robin scheduling across | |
1114 | * backend filesystem objects, there would be one resource associated with | |
1115 | * each of the backend filesystem objects partaking in the scheduling | |
1116 | * performed by the policy. | |
1117 | * | |
1118 | * NRS resources share a parent-child relationship, in which resources embedded | |
1119 | * in policy instances are the parent entities, with all scheduling entities | |
1120 | * a policy schedules across being the children, thus forming a simple resource | |
1121 | * hierarchy. This hierarchy may be extended with one or more levels in the | |
1122 | * future if the ability to have more than one primary policy is added. | |
1123 | * | |
1124 | * Upon request initialization, references to the then active NRS policies are | |
1125 | * taken and used to later handle the dispatching of the request with one of | |
1126 | * these policies. | |
1127 | * | |
1128 | * \see nrs_resource_get_safe() | |
1129 | * \see ptlrpc_nrs_req_add() | |
1130 | */ | |
1131 | struct ptlrpc_nrs_resource { | |
1132 | /** | |
1133 | * This NRS resource's parent; is NULL for resources embedded in NRS | |
1134 | * policy instances; i.e. those are top-level ones. | |
1135 | */ | |
1136 | struct ptlrpc_nrs_resource *res_parent; | |
1137 | /** | |
1138 | * The policy associated with this resource. | |
1139 | */ | |
1140 | struct ptlrpc_nrs_policy *res_policy; | |
1141 | }; | |
1142 | ||
1143 | enum { | |
1144 | NRS_RES_FALLBACK, | |
1145 | NRS_RES_PRIMARY, | |
1146 | NRS_RES_MAX | |
1147 | }; | |
1148 | ||
1149 | /* \name fifo | |
1150 | * | |
1151 | * FIFO policy | |
1152 | * | |
1153 | * This policy is a logical wrapper around previous, non-NRS functionality. | |
1154 | * It dispatches RPCs in the same order as they arrive from the network. This | |
1155 | * policy is currently used as the fallback policy, and the only enabled policy | |
1156 | * on all NRS heads of all PTLRPC service partitions. | |
1157 | * @{ | |
1158 | */ | |
1159 | ||
1160 | /** | |
1161 | * Private data structure for the FIFO policy | |
1162 | */ | |
1163 | struct nrs_fifo_head { | |
1164 | /** | |
1165 | * Resource object for policy instance. | |
1166 | */ | |
1167 | struct ptlrpc_nrs_resource fh_res; | |
1168 | /** | |
1169 | * List of queued requests. | |
1170 | */ | |
1171 | struct list_head fh_list; | |
1172 | /** | |
1173 | * For debugging purposes. | |
1174 | */ | |
1175 | __u64 fh_sequence; | |
1176 | }; | |
1177 | ||
1178 | struct nrs_fifo_req { | |
1179 | struct list_head fr_list; | |
1180 | __u64 fr_sequence; | |
1181 | }; | |
1182 | ||
1183 | /** @} fifo */ | |
1184 | ||
1185 | /** | |
1186 | * \name CRR-N | |
1187 | * | |
1188 | * CRR-N, Client Round Robin over NIDs | |
1189 | * @{ | |
1190 | */ | |
1191 | ||
1192 | /** | |
1193 | * private data structure for CRR-N NRS | |
1194 | */ | |
1195 | struct nrs_crrn_net { | |
1196 | struct ptlrpc_nrs_resource cn_res; | |
1197 | cfs_binheap_t *cn_binheap; | |
6da6eabe | 1198 | struct cfs_hash *cn_cli_hash; |
d7e09d03 PT |
1199 | /** |
1200 | * Used when a new scheduling round commences, in order to synchronize | |
1201 | * all clients with the new round number. | |
1202 | */ | |
1203 | __u64 cn_round; | |
1204 | /** | |
1205 | * Determines the relevant ordering amongst request batches within a | |
1206 | * scheduling round. | |
1207 | */ | |
1208 | __u64 cn_sequence; | |
1209 | /** | |
1210 | * Round Robin quantum; the maximum number of RPCs that each request | |
1211 | * batch for each client can have in a scheduling round. | |
1212 | */ | |
1213 | __u16 cn_quantum; | |
1214 | }; | |
1215 | ||
1216 | /** | |
1217 | * Object representing a client in CRR-N, as identified by its NID | |
1218 | */ | |
1219 | struct nrs_crrn_client { | |
1220 | struct ptlrpc_nrs_resource cc_res; | |
1221 | struct hlist_node cc_hnode; | |
1222 | lnet_nid_t cc_nid; | |
1223 | /** | |
1224 | * The round number against which this client is currently scheduling | |
1225 | * requests. | |
1226 | */ | |
1227 | __u64 cc_round; | |
1228 | /** | |
1229 | * The sequence number used for requests scheduled by this client during | |
1230 | * the current round number. | |
1231 | */ | |
1232 | __u64 cc_sequence; | |
1233 | atomic_t cc_ref; | |
1234 | /** | |
1235 | * Round Robin quantum; the maximum number of RPCs the client is allowed | |
1236 | * to schedule in a single batch of each round. | |
1237 | */ | |
1238 | __u16 cc_quantum; | |
1239 | /** | |
1240 | * # of pending requests for this client, on all existing rounds | |
1241 | */ | |
1242 | __u16 cc_active; | |
1243 | }; | |
1244 | ||
1245 | /** | |
1246 | * CRR-N NRS request definition | |
1247 | */ | |
1248 | struct nrs_crrn_req { | |
1249 | /** | |
1250 | * Round number for this request; shared with all other requests in the | |
1251 | * same batch. | |
1252 | */ | |
1253 | __u64 cr_round; | |
1254 | /** | |
1255 | * Sequence number for this request; shared with all other requests in | |
1256 | * the same batch. | |
1257 | */ | |
1258 | __u64 cr_sequence; | |
1259 | }; | |
1260 | ||
1261 | /** | |
1262 | * CRR-N policy operations. | |
1263 | */ | |
1264 | enum nrs_ctl_crr { | |
1265 | /** | |
1266 | * Read the RR quantum size of a CRR-N policy. | |
1267 | */ | |
1268 | NRS_CTL_CRRN_RD_QUANTUM = PTLRPC_NRS_CTL_1ST_POL_SPEC, | |
1269 | /** | |
1270 | * Write the RR quantum size of a CRR-N policy. | |
1271 | */ | |
1272 | NRS_CTL_CRRN_WR_QUANTUM, | |
1273 | }; | |
1274 | ||
1275 | /** @} CRR-N */ | |
1276 | ||
1277 | /** | |
1278 | * \name ORR/TRR | |
1279 | * | |
1280 | * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies | |
1281 | * @{ | |
1282 | */ | |
1283 | ||
1284 | /** | |
1285 | * Lower and upper byte offsets of a brw RPC | |
1286 | */ | |
1287 | struct nrs_orr_req_range { | |
1288 | __u64 or_start; | |
1289 | __u64 or_end; | |
1290 | }; | |
1291 | ||
1292 | /** | |
1293 | * RPC types supported by the ORR/TRR policies | |
1294 | */ | |
1295 | enum nrs_orr_supp { | |
1296 | NOS_OST_READ = (1 << 0), | |
1297 | NOS_OST_WRITE = (1 << 1), | |
1298 | NOS_OST_RW = (NOS_OST_READ | NOS_OST_WRITE), | |
1299 | /** | |
1300 | * Default value for policies. | |
1301 | */ | |
1302 | NOS_DFLT = NOS_OST_READ | |
1303 | }; | |
1304 | ||
1305 | /** | |
1306 | * As unique keys for grouping RPCs together, we use the object's OST FID for | |
1307 | * the ORR policy, and the OST index for the TRR policy. | |
1308 | * | |
1309 | * XXX: We waste some space for TRR policy instances by using a union, but it | |
1310 | * allows to consolidate some of the code between ORR and TRR, and these | |
1311 | * policies will probably eventually merge into one anyway. | |
1312 | */ | |
1313 | struct nrs_orr_key { | |
1314 | union { | |
1315 | /** object FID for ORR */ | |
1316 | struct lu_fid ok_fid; | |
1317 | /** OST index for TRR */ | |
1318 | __u32 ok_idx; | |
1319 | }; | |
1320 | }; | |
1321 | ||
1322 | /** | |
1323 | * The largest base string for unique hash/slab object names is | |
1324 | * "nrs_orr_reg_", so 13 characters. We add 3 to this to be used for the CPT | |
1325 | * id number, so this _should_ be more than enough for the maximum number of | |
1326 | * CPTs on any system. If it does happen that this statement is incorrect, | |
1327 | * nrs_orr_genobjname() will inevitably yield a non-unique name and cause | |
1328 | * kmem_cache_create() to complain (on Linux), so the erroneous situation | |
1329 | * will hopefully not go unnoticed. | |
1330 | */ | |
1331 | #define NRS_ORR_OBJ_NAME_MAX (sizeof("nrs_orr_reg_") + 3) | |
1332 | ||
1333 | /** | |
1334 | * private data structure for ORR and TRR NRS | |
1335 | */ | |
1336 | struct nrs_orr_data { | |
1337 | struct ptlrpc_nrs_resource od_res; | |
1338 | cfs_binheap_t *od_binheap; | |
6da6eabe | 1339 | struct cfs_hash *od_obj_hash; |
d7e09d03 PT |
1340 | struct kmem_cache *od_cache; |
1341 | /** | |
1342 | * Used when a new scheduling round commences, in order to synchronize | |
1343 | * all object or OST batches with the new round number. | |
1344 | */ | |
1345 | __u64 od_round; | |
1346 | /** | |
1347 | * Determines the relevant ordering amongst request batches within a | |
1348 | * scheduling round. | |
1349 | */ | |
1350 | __u64 od_sequence; | |
1351 | /** | |
1352 | * RPC types that are currently supported. | |
1353 | */ | |
1354 | enum nrs_orr_supp od_supp; | |
1355 | /** | |
17891183 | 1356 | * Round Robin quantum; the maximum number of RPCs that each request |
d7e09d03 PT |
1357 | * batch for each object or OST can have in a scheduling round. |
1358 | */ | |
1359 | __u16 od_quantum; | |
1360 | /** | |
1361 | * Whether to use physical disk offsets or logical file offsets. | |
1362 | */ | |
1363 | bool od_physical; | |
1364 | /** | |
1365 | * XXX: We need to provide a persistently allocated string to hold | |
1366 | * unique object names for this policy, since in currently supported | |
1367 | * versions of Linux by Lustre, kmem_cache_create() just sets a pointer | |
1368 | * to the name string provided. kstrdup() is used in the version of | |
1369 | * kmeme_cache_create() in current Linux mainline, so we may be able to | |
1370 | * remove this in the future. | |
1371 | */ | |
1372 | char od_objname[NRS_ORR_OBJ_NAME_MAX]; | |
1373 | }; | |
1374 | ||
1375 | /** | |
1376 | * Represents a backend-fs object or OST in the ORR and TRR policies | |
1377 | * respectively | |
1378 | */ | |
1379 | struct nrs_orr_object { | |
1380 | struct ptlrpc_nrs_resource oo_res; | |
1381 | struct hlist_node oo_hnode; | |
1382 | /** | |
1383 | * The round number against which requests are being scheduled for this | |
1384 | * object or OST | |
1385 | */ | |
1386 | __u64 oo_round; | |
1387 | /** | |
1388 | * The sequence number used for requests scheduled for this object or | |
1389 | * OST during the current round number. | |
1390 | */ | |
1391 | __u64 oo_sequence; | |
1392 | /** | |
1393 | * The key of the object or OST for which this structure instance is | |
1394 | * scheduling RPCs | |
1395 | */ | |
1396 | struct nrs_orr_key oo_key; | |
1397 | atomic_t oo_ref; | |
1398 | /** | |
1399 | * Round Robin quantum; the maximum number of RPCs that are allowed to | |
1400 | * be scheduled for the object or OST in a single batch of each round. | |
1401 | */ | |
1402 | __u16 oo_quantum; | |
1403 | /** | |
1404 | * # of pending requests for this object or OST, on all existing rounds | |
1405 | */ | |
1406 | __u16 oo_active; | |
1407 | }; | |
1408 | ||
1409 | /** | |
1410 | * ORR/TRR NRS request definition | |
1411 | */ | |
1412 | struct nrs_orr_req { | |
1413 | /** | |
1414 | * The offset range this request covers | |
1415 | */ | |
1416 | struct nrs_orr_req_range or_range; | |
1417 | /** | |
1418 | * Round number for this request; shared with all other requests in the | |
1419 | * same batch. | |
1420 | */ | |
1421 | __u64 or_round; | |
1422 | /** | |
1423 | * Sequence number for this request; shared with all other requests in | |
1424 | * the same batch. | |
1425 | */ | |
1426 | __u64 or_sequence; | |
1427 | /** | |
1428 | * For debugging purposes. | |
1429 | */ | |
1430 | struct nrs_orr_key or_key; | |
1431 | /** | |
1432 | * An ORR policy instance has filled in request information while | |
1433 | * enqueueing the request on the service partition's regular NRS head. | |
1434 | */ | |
1435 | unsigned int or_orr_set:1; | |
1436 | /** | |
1437 | * A TRR policy instance has filled in request information while | |
1438 | * enqueueing the request on the service partition's regular NRS head. | |
1439 | */ | |
1440 | unsigned int or_trr_set:1; | |
1441 | /** | |
1442 | * Request offset ranges have been filled in with logical offset | |
1443 | * values. | |
1444 | */ | |
1445 | unsigned int or_logical_set:1; | |
1446 | /** | |
1447 | * Request offset ranges have been filled in with physical offset | |
1448 | * values. | |
1449 | */ | |
1450 | unsigned int or_physical_set:1; | |
1451 | }; | |
1452 | ||
1453 | /** @} ORR/TRR */ | |
1454 | ||
1455 | /** | |
1456 | * NRS request | |
1457 | * | |
1458 | * Instances of this object exist embedded within ptlrpc_request; the main | |
1459 | * purpose of this object is to hold references to the request's resources | |
1460 | * for the lifetime of the request, and to hold properties that policies use | |
1461 | * use for determining the request's scheduling priority. | |
1462 | * */ | |
1463 | struct ptlrpc_nrs_request { | |
1464 | /** | |
1465 | * The request's resource hierarchy. | |
1466 | */ | |
1467 | struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX]; | |
1468 | /** | |
1469 | * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the | |
1470 | * policy that was used to enqueue the request. | |
1471 | * | |
1472 | * \see nrs_request_enqueue() | |
1473 | */ | |
1474 | unsigned nr_res_idx; | |
1475 | unsigned nr_initialized:1; | |
1476 | unsigned nr_enqueued:1; | |
1477 | unsigned nr_started:1; | |
1478 | unsigned nr_finalized:1; | |
1479 | cfs_binheap_node_t nr_node; | |
1480 | ||
1481 | /** | |
1482 | * Policy-specific fields, used for determining a request's scheduling | |
1483 | * priority, and other supporting functionality. | |
1484 | */ | |
1485 | union { | |
1486 | /** | |
1487 | * Fields for the FIFO policy | |
1488 | */ | |
1489 | struct nrs_fifo_req fifo; | |
1490 | /** | |
17891183 | 1491 | * CRR-N request definition |
d7e09d03 PT |
1492 | */ |
1493 | struct nrs_crrn_req crr; | |
1494 | /** ORR and TRR share the same request definition */ | |
1495 | struct nrs_orr_req orr; | |
1496 | } nr_u; | |
1497 | /** | |
1498 | * Externally-registering policies may want to use this to allocate | |
1499 | * their own request properties. | |
1500 | */ | |
1501 | void *ext; | |
1502 | }; | |
1503 | ||
1504 | /** @} nrs */ | |
1505 | ||
1506 | /** | |
1507 | * Basic request prioritization operations structure. | |
1508 | * The whole idea is centered around locks and RPCs that might affect locks. | |
1509 | * When a lock is contended we try to give priority to RPCs that might lead | |
1510 | * to fastest release of that lock. | |
1511 | * Currently only implemented for OSTs only in a way that makes all | |
1512 | * IO and truncate RPCs that are coming from a locked region where a lock is | |
1513 | * contended a priority over other requests. | |
1514 | */ | |
1515 | struct ptlrpc_hpreq_ops { | |
1516 | /** | |
1517 | * Check if the lock handle of the given lock is the same as | |
1518 | * taken from the request. | |
1519 | */ | |
1520 | int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *); | |
1521 | /** | |
1522 | * Check if the request is a high priority one. | |
1523 | */ | |
1524 | int (*hpreq_check)(struct ptlrpc_request *); | |
1525 | /** | |
1526 | * Called after the request has been handled. | |
1527 | */ | |
1528 | void (*hpreq_fini)(struct ptlrpc_request *); | |
1529 | }; | |
1530 | ||
1531 | /** | |
1532 | * Represents remote procedure call. | |
1533 | * | |
1534 | * This is a staple structure used by everybody wanting to send a request | |
1535 | * in Lustre. | |
1536 | */ | |
1537 | struct ptlrpc_request { | |
1538 | /* Request type: one of PTL_RPC_MSG_* */ | |
1539 | int rq_type; | |
1540 | /** Result of request processing */ | |
1541 | int rq_status; | |
1542 | /** | |
1543 | * Linkage item through which this request is included into | |
1544 | * sending/delayed lists on client and into rqbd list on server | |
1545 | */ | |
1546 | struct list_head rq_list; | |
1547 | /** | |
1548 | * Server side list of incoming unserved requests sorted by arrival | |
1549 | * time. Traversed from time to time to notice about to expire | |
1550 | * requests and sent back "early replies" to clients to let them | |
1551 | * know server is alive and well, just very busy to service their | |
1552 | * requests in time | |
1553 | */ | |
1554 | struct list_head rq_timed_list; | |
17891183 | 1555 | /** server-side history, used for debugging purposes. */ |
d7e09d03 PT |
1556 | struct list_head rq_history_list; |
1557 | /** server-side per-export list */ | |
1558 | struct list_head rq_exp_list; | |
1559 | /** server-side hp handlers */ | |
1560 | struct ptlrpc_hpreq_ops *rq_ops; | |
1561 | ||
1562 | /** initial thread servicing this request */ | |
1563 | struct ptlrpc_thread *rq_svc_thread; | |
1564 | ||
1565 | /** history sequence # */ | |
1566 | __u64 rq_history_seq; | |
1567 | /** \addtogroup nrs | |
1568 | * @{ | |
1569 | */ | |
1570 | /** stub for NRS request */ | |
1571 | struct ptlrpc_nrs_request rq_nrq; | |
1572 | /** @} nrs */ | |
1573 | /** the index of service's srv_at_array into which request is linked */ | |
1574 | time_t rq_at_index; | |
1575 | /** Lock to protect request flags and some other important bits, like | |
1576 | * rq_list | |
1577 | */ | |
1578 | spinlock_t rq_lock; | |
1579 | /** client-side flags are serialized by rq_lock */ | |
1580 | unsigned int rq_intr:1, rq_replied:1, rq_err:1, | |
1581 | rq_timedout:1, rq_resend:1, rq_restart:1, | |
1582 | /** | |
1583 | * when ->rq_replay is set, request is kept by the client even | |
1584 | * after server commits corresponding transaction. This is | |
1585 | * used for operations that require sequence of multiple | |
1586 | * requests to be replayed. The only example currently is file | |
1587 | * open/close. When last request in such a sequence is | |
1588 | * committed, ->rq_replay is cleared on all requests in the | |
1589 | * sequence. | |
1590 | */ | |
1591 | rq_replay:1, | |
1592 | rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1, | |
1593 | rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1, | |
cf378ff7 AL |
1594 | rq_early:1, |
1595 | rq_req_unlink:1, rq_reply_unlink:1, | |
d7e09d03 PT |
1596 | rq_memalloc:1, /* req originated from "kswapd" */ |
1597 | /* server-side flags */ | |
1598 | rq_packed_final:1, /* packed final reply */ | |
1599 | rq_hp:1, /* high priority RPC */ | |
1600 | rq_at_linked:1, /* link into service's srv_at_array */ | |
1601 | rq_reply_truncate:1, | |
1602 | rq_committed:1, | |
1603 | /* whether the "rq_set" is a valid one */ | |
1604 | rq_invalid_rqset:1, | |
1605 | rq_generation_set:1, | |
1606 | /* do not resend request on -EINPROGRESS */ | |
1607 | rq_no_retry_einprogress:1, | |
1608 | /* allow the req to be sent if the import is in recovery | |
1609 | * status */ | |
c52f69c5 | 1610 | rq_allow_replay:1; |
d7e09d03 PT |
1611 | |
1612 | unsigned int rq_nr_resend; | |
1613 | ||
1614 | enum rq_phase rq_phase; /* one of RQ_PHASE_* */ | |
1615 | enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */ | |
1616 | atomic_t rq_refcount;/* client-side refcount for SENT race, | |
17891183 | 1617 | server-side refcount for multiple replies */ |
d7e09d03 PT |
1618 | |
1619 | /** Portal to which this request would be sent */ | |
1620 | short rq_request_portal; /* XXX FIXME bug 249 */ | |
1621 | /** Portal where to wait for reply and where reply would be sent */ | |
1622 | short rq_reply_portal; /* XXX FIXME bug 249 */ | |
1623 | ||
1624 | /** | |
1625 | * client-side: | |
1626 | * !rq_truncate : # reply bytes actually received, | |
1627 | * rq_truncate : required repbuf_len for resend | |
1628 | */ | |
1629 | int rq_nob_received; | |
1630 | /** Request length */ | |
1631 | int rq_reqlen; | |
1632 | /** Reply length */ | |
1633 | int rq_replen; | |
1634 | /** Request message - what client sent */ | |
1635 | struct lustre_msg *rq_reqmsg; | |
1636 | /** Reply message - server response */ | |
1637 | struct lustre_msg *rq_repmsg; | |
1638 | /** Transaction number */ | |
1639 | __u64 rq_transno; | |
1640 | /** xid */ | |
1641 | __u64 rq_xid; | |
1642 | /** | |
17891183 | 1643 | * List item to for replay list. Not yet committed requests get linked |
d7e09d03 PT |
1644 | * there. |
1645 | * Also see \a rq_replay comment above. | |
1646 | */ | |
1647 | struct list_head rq_replay_list; | |
1648 | ||
1649 | /** | |
1650 | * security and encryption data | |
1651 | * @{ */ | |
1652 | struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */ | |
1653 | struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */ | |
1654 | struct list_head rq_ctx_chain; /**< link to waited ctx */ | |
1655 | ||
1656 | struct sptlrpc_flavor rq_flvr; /**< for client & server */ | |
1657 | enum lustre_sec_part rq_sp_from; | |
1658 | ||
1659 | /* client/server security flags */ | |
1660 | unsigned int | |
1661 | rq_ctx_init:1, /* context initiation */ | |
1662 | rq_ctx_fini:1, /* context destroy */ | |
1663 | rq_bulk_read:1, /* request bulk read */ | |
1664 | rq_bulk_write:1, /* request bulk write */ | |
1665 | /* server authentication flags */ | |
1666 | rq_auth_gss:1, /* authenticated by gss */ | |
1667 | rq_auth_remote:1, /* authed as remote user */ | |
1668 | rq_auth_usr_root:1, /* authed as root */ | |
1669 | rq_auth_usr_mdt:1, /* authed as mdt */ | |
1670 | rq_auth_usr_ost:1, /* authed as ost */ | |
1671 | /* security tfm flags */ | |
1672 | rq_pack_udesc:1, | |
1673 | rq_pack_bulk:1, | |
1674 | /* doesn't expect reply FIXME */ | |
1675 | rq_no_reply:1, | |
1676 | rq_pill_init:1; /* pill initialized */ | |
1677 | ||
1678 | uid_t rq_auth_uid; /* authed uid */ | |
1679 | uid_t rq_auth_mapped_uid; /* authed uid mapped to */ | |
1680 | ||
1681 | /* (server side), pointed directly into req buffer */ | |
1682 | struct ptlrpc_user_desc *rq_user_desc; | |
1683 | ||
1684 | /* various buffer pointers */ | |
1685 | struct lustre_msg *rq_reqbuf; /* req wrapper */ | |
1686 | char *rq_repbuf; /* rep buffer */ | |
1687 | struct lustre_msg *rq_repdata; /* rep wrapper msg */ | |
1688 | struct lustre_msg *rq_clrbuf; /* only in priv mode */ | |
1689 | int rq_reqbuf_len; /* req wrapper buf len */ | |
1690 | int rq_reqdata_len; /* req wrapper msg len */ | |
1691 | int rq_repbuf_len; /* rep buffer len */ | |
1692 | int rq_repdata_len; /* rep wrapper msg len */ | |
1693 | int rq_clrbuf_len; /* only in priv mode */ | |
1694 | int rq_clrdata_len; /* only in priv mode */ | |
1695 | ||
1696 | /** early replies go to offset 0, regular replies go after that */ | |
1697 | unsigned int rq_reply_off; | |
1698 | ||
1699 | /** @} */ | |
1700 | ||
1701 | /** Fields that help to see if request and reply were swabbed or not */ | |
1702 | __u32 rq_req_swab_mask; | |
1703 | __u32 rq_rep_swab_mask; | |
1704 | ||
1705 | /** What was import generation when this request was sent */ | |
1706 | int rq_import_generation; | |
1707 | enum lustre_imp_state rq_send_state; | |
1708 | ||
1709 | /** how many early replies (for stats) */ | |
1710 | int rq_early_count; | |
1711 | ||
1712 | /** client+server request */ | |
1713 | lnet_handle_md_t rq_req_md_h; | |
1714 | struct ptlrpc_cb_id rq_req_cbid; | |
1715 | /** optional time limit for send attempts */ | |
1716 | cfs_duration_t rq_delay_limit; | |
1717 | /** time request was first queued */ | |
a649ad1d | 1718 | unsigned long rq_queued_time; |
d7e09d03 PT |
1719 | |
1720 | /* server-side... */ | |
1721 | /** request arrival time */ | |
1722 | struct timeval rq_arrival_time; | |
1723 | /** separated reply state */ | |
1724 | struct ptlrpc_reply_state *rq_reply_state; | |
1725 | /** incoming request buffer */ | |
1726 | struct ptlrpc_request_buffer_desc *rq_rqbd; | |
1727 | ||
1728 | /** client-only incoming reply */ | |
1729 | lnet_handle_md_t rq_reply_md_h; | |
1730 | wait_queue_head_t rq_reply_waitq; | |
1731 | struct ptlrpc_cb_id rq_reply_cbid; | |
1732 | ||
1733 | /** our LNet NID */ | |
1734 | lnet_nid_t rq_self; | |
1735 | /** Peer description (the other side) */ | |
1736 | lnet_process_id_t rq_peer; | |
1737 | /** Server-side, export on which request was received */ | |
1738 | struct obd_export *rq_export; | |
1739 | /** Client side, import where request is being sent */ | |
1740 | struct obd_import *rq_import; | |
1741 | ||
1742 | /** Replay callback, called after request is replayed at recovery */ | |
1743 | void (*rq_replay_cb)(struct ptlrpc_request *); | |
1744 | /** | |
1745 | * Commit callback, called when request is committed and about to be | |
1746 | * freed. | |
1747 | */ | |
1748 | void (*rq_commit_cb)(struct ptlrpc_request *); | |
1749 | /** Opaq data for replay and commit callbacks. */ | |
1750 | void *rq_cb_data; | |
1751 | ||
1752 | /** For bulk requests on client only: bulk descriptor */ | |
1753 | struct ptlrpc_bulk_desc *rq_bulk; | |
1754 | ||
1755 | /** client outgoing req */ | |
1756 | /** | |
1757 | * when request/reply sent (secs), or time when request should be sent | |
1758 | */ | |
1759 | time_t rq_sent; | |
1760 | /** time for request really sent out */ | |
1761 | time_t rq_real_sent; | |
1762 | ||
1763 | /** when request must finish. volatile | |
1764 | * so that servers' early reply updates to the deadline aren't | |
1765 | * kept in per-cpu cache */ | |
1766 | volatile time_t rq_deadline; | |
1767 | /** when req reply unlink must finish. */ | |
1768 | time_t rq_reply_deadline; | |
1769 | /** when req bulk unlink must finish. */ | |
1770 | time_t rq_bulk_deadline; | |
1771 | /** | |
1772 | * service time estimate (secs) | |
1773 | * If the requestsis not served by this time, it is marked as timed out. | |
1774 | */ | |
1775 | int rq_timeout; | |
1776 | ||
1777 | /** Multi-rpc bits */ | |
1778 | /** Per-request waitq introduced by bug 21938 for recovery waiting */ | |
1779 | wait_queue_head_t rq_set_waitq; | |
1780 | /** Link item for request set lists */ | |
1781 | struct list_head rq_set_chain; | |
1782 | /** Link back to the request set */ | |
1783 | struct ptlrpc_request_set *rq_set; | |
1784 | /** Async completion handler, called when reply is received */ | |
1785 | ptlrpc_interpterer_t rq_interpret_reply; | |
1786 | /** Async completion context */ | |
1787 | union ptlrpc_async_args rq_async_args; | |
1788 | ||
1789 | /** Pool if request is from preallocated list */ | |
1790 | struct ptlrpc_request_pool *rq_pool; | |
1791 | ||
1792 | struct lu_context rq_session; | |
1793 | struct lu_context rq_recov_session; | |
1794 | ||
1795 | /** request format description */ | |
1796 | struct req_capsule rq_pill; | |
1797 | }; | |
1798 | ||
1799 | /** | |
1800 | * Call completion handler for rpc if any, return it's status or original | |
1801 | * rc if there was no handler defined for this request. | |
1802 | */ | |
1803 | static inline int ptlrpc_req_interpret(const struct lu_env *env, | |
1804 | struct ptlrpc_request *req, int rc) | |
1805 | { | |
1806 | if (req->rq_interpret_reply != NULL) { | |
1807 | req->rq_status = req->rq_interpret_reply(env, req, | |
1808 | &req->rq_async_args, | |
1809 | rc); | |
1810 | return req->rq_status; | |
1811 | } | |
1812 | return rc; | |
1813 | } | |
1814 | ||
1815 | /** \addtogroup nrs | |
1816 | * @{ | |
1817 | */ | |
1818 | int ptlrpc_nrs_policy_register(struct ptlrpc_nrs_pol_conf *conf); | |
1819 | int ptlrpc_nrs_policy_unregister(struct ptlrpc_nrs_pol_conf *conf); | |
1820 | void ptlrpc_nrs_req_hp_move(struct ptlrpc_request *req); | |
1821 | void nrs_policy_get_info_locked(struct ptlrpc_nrs_policy *policy, | |
1822 | struct ptlrpc_nrs_pol_info *info); | |
1823 | ||
1824 | /* | |
1825 | * Can the request be moved from the regular NRS head to the high-priority NRS | |
1826 | * head (of the same PTLRPC service partition), if any? | |
1827 | * | |
1828 | * For a reliable result, this should be checked under svcpt->scp_req lock. | |
1829 | */ | |
1830 | static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req) | |
1831 | { | |
1832 | struct ptlrpc_nrs_request *nrq = &req->rq_nrq; | |
1833 | ||
1834 | /** | |
1835 | * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the | |
1836 | * request has been enqueued first, and ptlrpc_nrs_request::nr_started | |
1837 | * to make sure it has not been scheduled yet (analogous to previous | |
1838 | * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list). | |
1839 | */ | |
1840 | return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp; | |
1841 | } | |
1842 | /** @} nrs */ | |
1843 | ||
1844 | /** | |
1845 | * Returns 1 if request buffer at offset \a index was already swabbed | |
1846 | */ | |
1847 | static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index) | |
1848 | { | |
1849 | LASSERT(index < sizeof(req->rq_req_swab_mask) * 8); | |
1850 | return req->rq_req_swab_mask & (1 << index); | |
1851 | } | |
1852 | ||
1853 | /** | |
1854 | * Returns 1 if request reply buffer at offset \a index was already swabbed | |
1855 | */ | |
1856 | static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index) | |
1857 | { | |
1858 | LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8); | |
1859 | return req->rq_rep_swab_mask & (1 << index); | |
1860 | } | |
1861 | ||
1862 | /** | |
1863 | * Returns 1 if request needs to be swabbed into local cpu byteorder | |
1864 | */ | |
1865 | static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req) | |
1866 | { | |
1867 | return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF); | |
1868 | } | |
1869 | ||
1870 | /** | |
1871 | * Returns 1 if request reply needs to be swabbed into local cpu byteorder | |
1872 | */ | |
1873 | static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req) | |
1874 | { | |
1875 | return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF); | |
1876 | } | |
1877 | ||
1878 | /** | |
1879 | * Mark request buffer at offset \a index that it was already swabbed | |
1880 | */ | |
1881 | static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index) | |
1882 | { | |
1883 | LASSERT(index < sizeof(req->rq_req_swab_mask) * 8); | |
1884 | LASSERT((req->rq_req_swab_mask & (1 << index)) == 0); | |
1885 | req->rq_req_swab_mask |= 1 << index; | |
1886 | } | |
1887 | ||
1888 | /** | |
1889 | * Mark request reply buffer at offset \a index that it was already swabbed | |
1890 | */ | |
1891 | static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index) | |
1892 | { | |
1893 | LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8); | |
1894 | LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0); | |
1895 | req->rq_rep_swab_mask |= 1 << index; | |
1896 | } | |
1897 | ||
1898 | /** | |
1899 | * Convert numerical request phase value \a phase into text string description | |
1900 | */ | |
1901 | static inline const char * | |
1902 | ptlrpc_phase2str(enum rq_phase phase) | |
1903 | { | |
1904 | switch (phase) { | |
1905 | case RQ_PHASE_NEW: | |
1906 | return "New"; | |
1907 | case RQ_PHASE_RPC: | |
1908 | return "Rpc"; | |
1909 | case RQ_PHASE_BULK: | |
1910 | return "Bulk"; | |
1911 | case RQ_PHASE_INTERPRET: | |
1912 | return "Interpret"; | |
1913 | case RQ_PHASE_COMPLETE: | |
1914 | return "Complete"; | |
1915 | case RQ_PHASE_UNREGISTERING: | |
1916 | return "Unregistering"; | |
1917 | default: | |
1918 | return "?Phase?"; | |
1919 | } | |
1920 | } | |
1921 | ||
1922 | /** | |
1923 | * Convert numerical request phase of the request \a req into text stringi | |
1924 | * description | |
1925 | */ | |
1926 | static inline const char * | |
1927 | ptlrpc_rqphase2str(struct ptlrpc_request *req) | |
1928 | { | |
1929 | return ptlrpc_phase2str(req->rq_phase); | |
1930 | } | |
1931 | ||
1932 | /** | |
1933 | * Debugging functions and helpers to print request structure into debug log | |
1934 | * @{ | |
1935 | */ | |
1936 | /* Spare the preprocessor, spoil the bugs. */ | |
1937 | #define FLAG(field, str) (field ? str : "") | |
1938 | ||
1939 | /** Convert bit flags into a string */ | |
1940 | #define DEBUG_REQ_FLAGS(req) \ | |
1941 | ptlrpc_rqphase2str(req), \ | |
1942 | FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \ | |
1943 | FLAG(req->rq_err, "E"), \ | |
1944 | FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \ | |
1945 | FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \ | |
1946 | FLAG(req->rq_no_resend, "N"), \ | |
1947 | FLAG(req->rq_waiting, "W"), \ | |
1948 | FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \ | |
1949 | FLAG(req->rq_committed, "M") | |
1950 | ||
1951 | #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s" | |
1952 | ||
1953 | void _debug_req(struct ptlrpc_request *req, | |
1954 | struct libcfs_debug_msg_data *data, const char *fmt, ...) | |
1955 | __attribute__ ((format (printf, 3, 4))); | |
1956 | ||
1957 | /** | |
17891183 | 1958 | * Helper that decides if we need to print request according to current debug |
d7e09d03 PT |
1959 | * level settings |
1960 | */ | |
1961 | #define debug_req(msgdata, mask, cdls, req, fmt, a...) \ | |
1962 | do { \ | |
1963 | CFS_CHECK_STACK(msgdata, mask, cdls); \ | |
1964 | \ | |
1965 | if (((mask) & D_CANTMASK) != 0 || \ | |
1966 | ((libcfs_debug & (mask)) != 0 && \ | |
1967 | (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \ | |
1968 | _debug_req((req), msgdata, fmt, ##a); \ | |
1969 | } while(0) | |
1970 | ||
1971 | /** | |
17891183 | 1972 | * This is the debug print function you need to use to print request structure |
d7e09d03 PT |
1973 | * content into lustre debug log. |
1974 | * for most callers (level is a constant) this is resolved at compile time */ | |
1975 | #define DEBUG_REQ(level, req, fmt, args...) \ | |
1976 | do { \ | |
1977 | if ((level) & (D_ERROR | D_WARNING)) { \ | |
a3ea59e0 | 1978 | static struct cfs_debug_limit_state cdls; \ |
d7e09d03 PT |
1979 | LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \ |
1980 | debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\ | |
1981 | } else { \ | |
1982 | LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \ | |
1983 | debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \ | |
1984 | } \ | |
1985 | } while (0) | |
1986 | /** @} */ | |
1987 | ||
1988 | /** | |
1989 | * Structure that defines a single page of a bulk transfer | |
1990 | */ | |
1991 | struct ptlrpc_bulk_page { | |
1992 | /** Linkage to list of pages in a bulk */ | |
1993 | struct list_head bp_link; | |
1994 | /** | |
1995 | * Number of bytes in a page to transfer starting from \a bp_pageoffset | |
1996 | */ | |
1997 | int bp_buflen; | |
1998 | /** offset within a page */ | |
1999 | int bp_pageoffset; | |
2000 | /** The page itself */ | |
2001 | struct page *bp_page; | |
2002 | }; | |
2003 | ||
2004 | #define BULK_GET_SOURCE 0 | |
2005 | #define BULK_PUT_SINK 1 | |
2006 | #define BULK_GET_SINK 2 | |
2007 | #define BULK_PUT_SOURCE 3 | |
2008 | ||
2009 | /** | |
2010 | * Definition of bulk descriptor. | |
2011 | * Bulks are special "Two phase" RPCs where initial request message | |
2012 | * is sent first and it is followed bt a transfer (o receiving) of a large | |
2013 | * amount of data to be settled into pages referenced from the bulk descriptors. | |
2014 | * Bulks transfers (the actual data following the small requests) are done | |
2015 | * on separate LNet portals. | |
2016 | * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs. | |
2017 | * Another user is readpage for MDT. | |
2018 | */ | |
2019 | struct ptlrpc_bulk_desc { | |
2020 | /** completed with failure */ | |
2021 | unsigned long bd_failure:1; | |
2022 | /** {put,get}{source,sink} */ | |
2023 | unsigned long bd_type:2; | |
2024 | /** client side */ | |
2025 | unsigned long bd_registered:1; | |
2026 | /** For serialization with callback */ | |
2027 | spinlock_t bd_lock; | |
2028 | /** Import generation when request for this bulk was sent */ | |
2029 | int bd_import_generation; | |
2030 | /** LNet portal for this bulk */ | |
2031 | __u32 bd_portal; | |
2032 | /** Server side - export this bulk created for */ | |
2033 | struct obd_export *bd_export; | |
2034 | /** Client side - import this bulk was sent on */ | |
2035 | struct obd_import *bd_import; | |
2036 | /** Back pointer to the request */ | |
2037 | struct ptlrpc_request *bd_req; | |
2038 | wait_queue_head_t bd_waitq; /* server side only WQ */ | |
2039 | int bd_iov_count; /* # entries in bd_iov */ | |
2040 | int bd_max_iov; /* allocated size of bd_iov */ | |
2041 | int bd_nob; /* # bytes covered */ | |
2042 | int bd_nob_transferred; /* # bytes GOT/PUT */ | |
2043 | ||
2044 | __u64 bd_last_xid; | |
2045 | ||
2046 | struct ptlrpc_cb_id bd_cbid; /* network callback info */ | |
2047 | lnet_nid_t bd_sender; /* stash event::sender */ | |
2048 | int bd_md_count; /* # valid entries in bd_mds */ | |
2049 | int bd_md_max_brw; /* max entries in bd_mds */ | |
2050 | /** array of associated MDs */ | |
2051 | lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT]; | |
2052 | ||
2053 | /* | |
2054 | * encrypt iov, size is either 0 or bd_iov_count. | |
2055 | */ | |
2056 | lnet_kiov_t *bd_enc_iov; | |
2057 | ||
2058 | lnet_kiov_t bd_iov[0]; | |
2059 | }; | |
2060 | ||
2061 | enum { | |
2062 | SVC_STOPPED = 1 << 0, | |
2063 | SVC_STOPPING = 1 << 1, | |
2064 | SVC_STARTING = 1 << 2, | |
2065 | SVC_RUNNING = 1 << 3, | |
2066 | SVC_EVENT = 1 << 4, | |
2067 | SVC_SIGNAL = 1 << 5, | |
2068 | }; | |
2069 | ||
2070 | #define PTLRPC_THR_NAME_LEN 32 | |
2071 | /** | |
2072 | * Definition of server service thread structure | |
2073 | */ | |
2074 | struct ptlrpc_thread { | |
2075 | /** | |
2076 | * List of active threads in svc->srv_threads | |
2077 | */ | |
2078 | struct list_head t_link; | |
2079 | /** | |
2080 | * thread-private data (preallocated memory) | |
2081 | */ | |
2082 | void *t_data; | |
2083 | __u32 t_flags; | |
2084 | /** | |
2085 | * service thread index, from ptlrpc_start_threads | |
2086 | */ | |
2087 | unsigned int t_id; | |
2088 | /** | |
2089 | * service thread pid | |
2090 | */ | |
2091 | pid_t t_pid; | |
2092 | /** | |
2093 | * put watchdog in the structure per thread b=14840 | |
5d4450c4 PT |
2094 | * |
2095 | * Lustre watchdog is removed for client in the hope | |
2096 | * of a generic watchdog can be merged in kernel. | |
2097 | * When that happens, we should add below back. | |
2098 | * | |
2099 | * struct lc_watchdog *t_watchdog; | |
d7e09d03 | 2100 | */ |
d7e09d03 PT |
2101 | /** |
2102 | * the svc this thread belonged to b=18582 | |
2103 | */ | |
2104 | struct ptlrpc_service_part *t_svcpt; | |
2105 | wait_queue_head_t t_ctl_waitq; | |
2106 | struct lu_env *t_env; | |
2107 | char t_name[PTLRPC_THR_NAME_LEN]; | |
2108 | }; | |
2109 | ||
2110 | static inline int thread_is_init(struct ptlrpc_thread *thread) | |
2111 | { | |
2112 | return thread->t_flags == 0; | |
2113 | } | |
2114 | ||
2115 | static inline int thread_is_stopped(struct ptlrpc_thread *thread) | |
2116 | { | |
2117 | return !!(thread->t_flags & SVC_STOPPED); | |
2118 | } | |
2119 | ||
2120 | static inline int thread_is_stopping(struct ptlrpc_thread *thread) | |
2121 | { | |
2122 | return !!(thread->t_flags & SVC_STOPPING); | |
2123 | } | |
2124 | ||
2125 | static inline int thread_is_starting(struct ptlrpc_thread *thread) | |
2126 | { | |
2127 | return !!(thread->t_flags & SVC_STARTING); | |
2128 | } | |
2129 | ||
2130 | static inline int thread_is_running(struct ptlrpc_thread *thread) | |
2131 | { | |
2132 | return !!(thread->t_flags & SVC_RUNNING); | |
2133 | } | |
2134 | ||
2135 | static inline int thread_is_event(struct ptlrpc_thread *thread) | |
2136 | { | |
2137 | return !!(thread->t_flags & SVC_EVENT); | |
2138 | } | |
2139 | ||
2140 | static inline int thread_is_signal(struct ptlrpc_thread *thread) | |
2141 | { | |
2142 | return !!(thread->t_flags & SVC_SIGNAL); | |
2143 | } | |
2144 | ||
2145 | static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags) | |
2146 | { | |
2147 | thread->t_flags &= ~flags; | |
2148 | } | |
2149 | ||
2150 | static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags) | |
2151 | { | |
2152 | thread->t_flags = flags; | |
2153 | } | |
2154 | ||
2155 | static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags) | |
2156 | { | |
2157 | thread->t_flags |= flags; | |
2158 | } | |
2159 | ||
2160 | static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread, | |
2161 | __u32 flags) | |
2162 | { | |
2163 | if (thread->t_flags & flags) { | |
2164 | thread->t_flags &= ~flags; | |
2165 | return 1; | |
2166 | } | |
2167 | return 0; | |
2168 | } | |
2169 | ||
2170 | /** | |
2171 | * Request buffer descriptor structure. | |
2172 | * This is a structure that contains one posted request buffer for service. | |
2173 | * Once data land into a buffer, event callback creates actual request and | |
2174 | * notifies wakes one of the service threads to process new incoming request. | |
2175 | * More than one request can fit into the buffer. | |
2176 | */ | |
2177 | struct ptlrpc_request_buffer_desc { | |
2178 | /** Link item for rqbds on a service */ | |
2179 | struct list_head rqbd_list; | |
2180 | /** History of requests for this buffer */ | |
2181 | struct list_head rqbd_reqs; | |
2182 | /** Back pointer to service for which this buffer is registered */ | |
2183 | struct ptlrpc_service_part *rqbd_svcpt; | |
2184 | /** LNet descriptor */ | |
2185 | lnet_handle_md_t rqbd_md_h; | |
2186 | int rqbd_refcount; | |
2187 | /** The buffer itself */ | |
2188 | char *rqbd_buffer; | |
2189 | struct ptlrpc_cb_id rqbd_cbid; | |
2190 | /** | |
2191 | * This "embedded" request structure is only used for the | |
2192 | * last request to fit into the buffer | |
2193 | */ | |
2194 | struct ptlrpc_request rqbd_req; | |
2195 | }; | |
2196 | ||
2197 | typedef int (*svc_handler_t)(struct ptlrpc_request *req); | |
2198 | ||
2199 | struct ptlrpc_service_ops { | |
2200 | /** | |
2201 | * if non-NULL called during thread creation (ptlrpc_start_thread()) | |
2202 | * to initialize service specific per-thread state. | |
2203 | */ | |
2204 | int (*so_thr_init)(struct ptlrpc_thread *thr); | |
2205 | /** | |
2206 | * if non-NULL called during thread shutdown (ptlrpc_main()) to | |
2207 | * destruct state created by ->srv_init(). | |
2208 | */ | |
2209 | void (*so_thr_done)(struct ptlrpc_thread *thr); | |
2210 | /** | |
2211 | * Handler function for incoming requests for this service | |
2212 | */ | |
2213 | int (*so_req_handler)(struct ptlrpc_request *req); | |
2214 | /** | |
2215 | * function to determine priority of the request, it's called | |
2216 | * on every new request | |
2217 | */ | |
2218 | int (*so_hpreq_handler)(struct ptlrpc_request *); | |
2219 | /** | |
2220 | * service-specific print fn | |
2221 | */ | |
2222 | void (*so_req_printer)(void *, struct ptlrpc_request *); | |
2223 | }; | |
2224 | ||
2225 | #ifndef __cfs_cacheline_aligned | |
2226 | /* NB: put it here for reducing patche dependence */ | |
2227 | # define __cfs_cacheline_aligned | |
2228 | #endif | |
2229 | ||
2230 | /** | |
2231 | * How many high priority requests to serve before serving one normal | |
2232 | * priority request | |
2233 | */ | |
2234 | #define PTLRPC_SVC_HP_RATIO 10 | |
2235 | ||
2236 | /** | |
2237 | * Definition of PortalRPC service. | |
2238 | * The service is listening on a particular portal (like tcp port) | |
2239 | * and perform actions for a specific server like IO service for OST | |
2240 | * or general metadata service for MDS. | |
2241 | */ | |
2242 | struct ptlrpc_service { | |
2243 | /** serialize /proc operations */ | |
2244 | spinlock_t srv_lock; | |
2245 | /** most often accessed fields */ | |
2246 | /** chain thru all services */ | |
2247 | struct list_head srv_list; | |
2248 | /** service operations table */ | |
2249 | struct ptlrpc_service_ops srv_ops; | |
2250 | /** only statically allocated strings here; we don't clean them */ | |
2251 | char *srv_name; | |
2252 | /** only statically allocated strings here; we don't clean them */ | |
2253 | char *srv_thread_name; | |
2254 | /** service thread list */ | |
2255 | struct list_head srv_threads; | |
2256 | /** threads # should be created for each partition on initializing */ | |
2257 | int srv_nthrs_cpt_init; | |
2258 | /** limit of threads number for each partition */ | |
2259 | int srv_nthrs_cpt_limit; | |
2260 | /** Root of /proc dir tree for this service */ | |
b59fe845 | 2261 | struct proc_dir_entry *srv_procroot; |
d7e09d03 PT |
2262 | /** Pointer to statistic data for this service */ |
2263 | struct lprocfs_stats *srv_stats; | |
2264 | /** # hp per lp reqs to handle */ | |
2265 | int srv_hpreq_ratio; | |
2266 | /** biggest request to receive */ | |
2267 | int srv_max_req_size; | |
2268 | /** biggest reply to send */ | |
2269 | int srv_max_reply_size; | |
2270 | /** size of individual buffers */ | |
2271 | int srv_buf_size; | |
2272 | /** # buffers to allocate in 1 group */ | |
2273 | int srv_nbuf_per_group; | |
2274 | /** Local portal on which to receive requests */ | |
2275 | __u32 srv_req_portal; | |
2276 | /** Portal on the client to send replies to */ | |
2277 | __u32 srv_rep_portal; | |
2278 | /** | |
2279 | * Tags for lu_context associated with this thread, see struct | |
2280 | * lu_context. | |
2281 | */ | |
2282 | __u32 srv_ctx_tags; | |
2283 | /** soft watchdog timeout multiplier */ | |
2284 | int srv_watchdog_factor; | |
2285 | /** under unregister_service */ | |
2286 | unsigned srv_is_stopping:1; | |
2287 | ||
2288 | /** max # request buffers in history per partition */ | |
2289 | int srv_hist_nrqbds_cpt_max; | |
2290 | /** number of CPTs this service bound on */ | |
2291 | int srv_ncpts; | |
2292 | /** CPTs array this service bound on */ | |
2293 | __u32 *srv_cpts; | |
2294 | /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */ | |
2295 | int srv_cpt_bits; | |
2296 | /** CPT table this service is running over */ | |
2297 | struct cfs_cpt_table *srv_cptable; | |
2298 | /** | |
2299 | * partition data for ptlrpc service | |
2300 | */ | |
2301 | struct ptlrpc_service_part *srv_parts[0]; | |
2302 | }; | |
2303 | ||
2304 | /** | |
2305 | * Definition of PortalRPC service partition data. | |
2306 | * Although a service only has one instance of it right now, but we | |
2307 | * will have multiple instances very soon (instance per CPT). | |
2308 | * | |
2309 | * it has four locks: | |
2310 | * \a scp_lock | |
2311 | * serialize operations on rqbd and requests waiting for preprocess | |
2312 | * \a scp_req_lock | |
2313 | * serialize operations active requests sent to this portal | |
2314 | * \a scp_at_lock | |
2315 | * serialize adaptive timeout stuff | |
2316 | * \a scp_rep_lock | |
2317 | * serialize operations on RS list (reply states) | |
2318 | * | |
2319 | * We don't have any use-case to take two or more locks at the same time | |
2320 | * for now, so there is no lock order issue. | |
2321 | */ | |
2322 | struct ptlrpc_service_part { | |
2323 | /** back reference to owner */ | |
2324 | struct ptlrpc_service *scp_service __cfs_cacheline_aligned; | |
2325 | /* CPT id, reserved */ | |
2326 | int scp_cpt; | |
2327 | /** always increasing number */ | |
2328 | int scp_thr_nextid; | |
2329 | /** # of starting threads */ | |
2330 | int scp_nthrs_starting; | |
2331 | /** # of stopping threads, reserved for shrinking threads */ | |
2332 | int scp_nthrs_stopping; | |
2333 | /** # running threads */ | |
2334 | int scp_nthrs_running; | |
2335 | /** service threads list */ | |
2336 | struct list_head scp_threads; | |
2337 | ||
2338 | /** | |
2339 | * serialize the following fields, used for protecting | |
2340 | * rqbd list and incoming requests waiting for preprocess, | |
2341 | * threads starting & stopping are also protected by this lock. | |
2342 | */ | |
2343 | spinlock_t scp_lock __cfs_cacheline_aligned; | |
2344 | /** total # req buffer descs allocated */ | |
2345 | int scp_nrqbds_total; | |
2346 | /** # posted request buffers for receiving */ | |
2347 | int scp_nrqbds_posted; | |
2348 | /** in progress of allocating rqbd */ | |
2349 | int scp_rqbd_allocating; | |
2350 | /** # incoming reqs */ | |
2351 | int scp_nreqs_incoming; | |
2352 | /** request buffers to be reposted */ | |
2353 | struct list_head scp_rqbd_idle; | |
2354 | /** req buffers receiving */ | |
2355 | struct list_head scp_rqbd_posted; | |
2356 | /** incoming reqs */ | |
2357 | struct list_head scp_req_incoming; | |
2358 | /** timeout before re-posting reqs, in tick */ | |
2359 | cfs_duration_t scp_rqbd_timeout; | |
2360 | /** | |
2361 | * all threads sleep on this. This wait-queue is signalled when new | |
2362 | * incoming request arrives and when difficult reply has to be handled. | |
2363 | */ | |
2364 | wait_queue_head_t scp_waitq; | |
2365 | ||
2366 | /** request history */ | |
2367 | struct list_head scp_hist_reqs; | |
2368 | /** request buffer history */ | |
2369 | struct list_head scp_hist_rqbds; | |
2370 | /** # request buffers in history */ | |
2371 | int scp_hist_nrqbds; | |
2372 | /** sequence number for request */ | |
2373 | __u64 scp_hist_seq; | |
2374 | /** highest seq culled from history */ | |
2375 | __u64 scp_hist_seq_culled; | |
2376 | ||
2377 | /** | |
2378 | * serialize the following fields, used for processing requests | |
2379 | * sent to this portal | |
2380 | */ | |
2381 | spinlock_t scp_req_lock __cfs_cacheline_aligned; | |
2382 | /** # reqs in either of the NRS heads below */ | |
2383 | /** # reqs being served */ | |
2384 | int scp_nreqs_active; | |
2385 | /** # HPreqs being served */ | |
2386 | int scp_nhreqs_active; | |
2387 | /** # hp requests handled */ | |
2388 | int scp_hreq_count; | |
2389 | ||
2390 | /** NRS head for regular requests */ | |
2391 | struct ptlrpc_nrs scp_nrs_reg; | |
2392 | /** NRS head for HP requests; this is only valid for services that can | |
2393 | * handle HP requests */ | |
2394 | struct ptlrpc_nrs *scp_nrs_hp; | |
2395 | ||
2396 | /** AT stuff */ | |
2397 | /** @{ */ | |
2398 | /** | |
2399 | * serialize the following fields, used for changes on | |
2400 | * adaptive timeout | |
2401 | */ | |
2402 | spinlock_t scp_at_lock __cfs_cacheline_aligned; | |
2403 | /** estimated rpc service time */ | |
2404 | struct adaptive_timeout scp_at_estimate; | |
2405 | /** reqs waiting for replies */ | |
2406 | struct ptlrpc_at_array scp_at_array; | |
2407 | /** early reply timer */ | |
54319351 | 2408 | struct timer_list scp_at_timer; |
d7e09d03 | 2409 | /** debug */ |
a649ad1d | 2410 | unsigned long scp_at_checktime; |
d7e09d03 PT |
2411 | /** check early replies */ |
2412 | unsigned scp_at_check; | |
2413 | /** @} */ | |
2414 | ||
2415 | /** | |
2416 | * serialize the following fields, used for processing | |
2417 | * replies for this portal | |
2418 | */ | |
2419 | spinlock_t scp_rep_lock __cfs_cacheline_aligned; | |
2420 | /** all the active replies */ | |
2421 | struct list_head scp_rep_active; | |
2422 | /** List of free reply_states */ | |
2423 | struct list_head scp_rep_idle; | |
2424 | /** waitq to run, when adding stuff to srv_free_rs_list */ | |
2425 | wait_queue_head_t scp_rep_waitq; | |
2426 | /** # 'difficult' replies */ | |
2427 | atomic_t scp_nreps_difficult; | |
2428 | }; | |
2429 | ||
2430 | #define ptlrpc_service_for_each_part(part, i, svc) \ | |
2431 | for (i = 0; \ | |
2432 | i < (svc)->srv_ncpts && \ | |
2433 | (svc)->srv_parts != NULL && \ | |
2434 | ((part) = (svc)->srv_parts[i]) != NULL; i++) | |
2435 | ||
2436 | /** | |
2437 | * Declaration of ptlrpcd control structure | |
2438 | */ | |
2439 | struct ptlrpcd_ctl { | |
2440 | /** | |
2441 | * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE) | |
2442 | */ | |
2443 | unsigned long pc_flags; | |
2444 | /** | |
2445 | * Thread lock protecting structure fields. | |
2446 | */ | |
2447 | spinlock_t pc_lock; | |
2448 | /** | |
2449 | * Start completion. | |
2450 | */ | |
2451 | struct completion pc_starting; | |
2452 | /** | |
2453 | * Stop completion. | |
2454 | */ | |
2455 | struct completion pc_finishing; | |
2456 | /** | |
2457 | * Thread requests set. | |
2458 | */ | |
2459 | struct ptlrpc_request_set *pc_set; | |
2460 | /** | |
2461 | * Thread name used in cfs_daemonize() | |
2462 | */ | |
2463 | char pc_name[16]; | |
2464 | /** | |
2465 | * Environment for request interpreters to run in. | |
2466 | */ | |
2467 | struct lu_env pc_env; | |
2468 | /** | |
2469 | * Index of ptlrpcd thread in the array. | |
2470 | */ | |
2471 | int pc_index; | |
2472 | /** | |
2473 | * Number of the ptlrpcd's partners. | |
2474 | */ | |
2475 | int pc_npartners; | |
2476 | /** | |
2477 | * Pointer to the array of partners' ptlrpcd_ctl structure. | |
2478 | */ | |
2479 | struct ptlrpcd_ctl **pc_partners; | |
2480 | /** | |
2481 | * Record the partner index to be processed next. | |
2482 | */ | |
2483 | int pc_cursor; | |
2484 | }; | |
2485 | ||
2486 | /* Bits for pc_flags */ | |
2487 | enum ptlrpcd_ctl_flags { | |
2488 | /** | |
2489 | * Ptlrpc thread start flag. | |
2490 | */ | |
2491 | LIOD_START = 1 << 0, | |
2492 | /** | |
2493 | * Ptlrpc thread stop flag. | |
2494 | */ | |
2495 | LIOD_STOP = 1 << 1, | |
2496 | /** | |
2497 | * Ptlrpc thread force flag (only stop force so far). | |
2498 | * This will cause aborting any inflight rpcs handled | |
2499 | * by thread if LIOD_STOP is specified. | |
2500 | */ | |
2501 | LIOD_FORCE = 1 << 2, | |
2502 | /** | |
2503 | * This is a recovery ptlrpc thread. | |
2504 | */ | |
2505 | LIOD_RECOVERY = 1 << 3, | |
2506 | /** | |
2507 | * The ptlrpcd is bound to some CPU core. | |
2508 | */ | |
2509 | LIOD_BIND = 1 << 4, | |
2510 | }; | |
2511 | ||
2512 | /** | |
2513 | * \addtogroup nrs | |
2514 | * @{ | |
2515 | * | |
2516 | * Service compatibility function; the policy is compatible with all services. | |
2517 | * | |
2518 | * \param[in] svc The service the policy is attempting to register with. | |
2519 | * \param[in] desc The policy descriptor | |
2520 | * | |
2521 | * \retval true The policy is compatible with the service | |
2522 | * | |
2523 | * \see ptlrpc_nrs_pol_desc::pd_compat() | |
2524 | */ | |
2525 | static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc, | |
2526 | const struct ptlrpc_nrs_pol_desc *desc) | |
2527 | { | |
2528 | return true; | |
2529 | } | |
2530 | ||
2531 | /** | |
2532 | * Service compatibility function; the policy is compatible with only a specific | |
2533 | * service which is identified by its human-readable name at | |
2534 | * ptlrpc_service::srv_name. | |
2535 | * | |
2536 | * \param[in] svc The service the policy is attempting to register with. | |
2537 | * \param[in] desc The policy descriptor | |
2538 | * | |
2539 | * \retval false The policy is not compatible with the service | |
2540 | * \retval true The policy is compatible with the service | |
2541 | * | |
2542 | * \see ptlrpc_nrs_pol_desc::pd_compat() | |
2543 | */ | |
2544 | static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc, | |
2545 | const struct ptlrpc_nrs_pol_desc *desc) | |
2546 | { | |
2547 | LASSERT(desc->pd_compat_svc_name != NULL); | |
2548 | return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0; | |
2549 | } | |
2550 | ||
2551 | /** @} nrs */ | |
2552 | ||
2553 | /* ptlrpc/events.c */ | |
2554 | extern lnet_handle_eq_t ptlrpc_eq_h; | |
2555 | extern int ptlrpc_uuid_to_peer(struct obd_uuid *uuid, | |
2556 | lnet_process_id_t *peer, lnet_nid_t *self); | |
2557 | /** | |
2558 | * These callbacks are invoked by LNet when something happened to | |
2559 | * underlying buffer | |
2560 | * @{ | |
2561 | */ | |
2562 | extern void request_out_callback(lnet_event_t *ev); | |
2563 | extern void reply_in_callback(lnet_event_t *ev); | |
2564 | extern void client_bulk_callback(lnet_event_t *ev); | |
2565 | extern void request_in_callback(lnet_event_t *ev); | |
2566 | extern void reply_out_callback(lnet_event_t *ev); | |
2567 | /** @} */ | |
2568 | ||
2569 | /* ptlrpc/connection.c */ | |
2570 | struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer, | |
2571 | lnet_nid_t self, | |
2572 | struct obd_uuid *uuid); | |
2573 | int ptlrpc_connection_put(struct ptlrpc_connection *c); | |
2574 | struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *); | |
2575 | int ptlrpc_connection_init(void); | |
2576 | void ptlrpc_connection_fini(void); | |
2577 | extern lnet_pid_t ptl_get_pid(void); | |
2578 | ||
2579 | /* ptlrpc/niobuf.c */ | |
2580 | /** | |
2581 | * Actual interfacing with LNet to put/get/register/unregister stuff | |
2582 | * @{ | |
2583 | */ | |
2584 | ||
2585 | int ptlrpc_register_bulk(struct ptlrpc_request *req); | |
2586 | int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async); | |
2587 | ||
2588 | static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req) | |
2589 | { | |
2590 | struct ptlrpc_bulk_desc *desc; | |
2591 | int rc; | |
2592 | ||
2593 | LASSERT(req != NULL); | |
2594 | desc = req->rq_bulk; | |
2595 | ||
2596 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) && | |
7264b8a5 | 2597 | req->rq_bulk_deadline > get_seconds()) |
d7e09d03 PT |
2598 | return 1; |
2599 | ||
2600 | if (!desc) | |
2601 | return 0; | |
2602 | ||
2603 | spin_lock(&desc->bd_lock); | |
2604 | rc = desc->bd_md_count; | |
2605 | spin_unlock(&desc->bd_lock); | |
2606 | return rc; | |
2607 | } | |
2608 | ||
2609 | #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01 | |
2610 | #define PTLRPC_REPLY_EARLY 0x02 | |
2611 | int ptlrpc_send_reply(struct ptlrpc_request *req, int flags); | |
2612 | int ptlrpc_reply(struct ptlrpc_request *req); | |
2613 | int ptlrpc_send_error(struct ptlrpc_request *req, int difficult); | |
2614 | int ptlrpc_error(struct ptlrpc_request *req); | |
2615 | void ptlrpc_resend_req(struct ptlrpc_request *request); | |
2616 | int ptlrpc_at_get_net_latency(struct ptlrpc_request *req); | |
2617 | int ptl_send_rpc(struct ptlrpc_request *request, int noreply); | |
2618 | int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd); | |
2619 | /** @} */ | |
2620 | ||
2621 | /* ptlrpc/client.c */ | |
2622 | /** | |
2623 | * Client-side portals API. Everything to send requests, receive replies, | |
2624 | * request queues, request management, etc. | |
2625 | * @{ | |
2626 | */ | |
63d42578 HZ |
2627 | void ptlrpc_request_committed(struct ptlrpc_request *req, int force); |
2628 | ||
d7e09d03 PT |
2629 | void ptlrpc_init_client(int req_portal, int rep_portal, char *name, |
2630 | struct ptlrpc_client *); | |
2631 | void ptlrpc_cleanup_client(struct obd_import *imp); | |
2632 | struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid); | |
2633 | ||
2634 | int ptlrpc_queue_wait(struct ptlrpc_request *req); | |
2635 | int ptlrpc_replay_req(struct ptlrpc_request *req); | |
2636 | int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async); | |
2637 | void ptlrpc_restart_req(struct ptlrpc_request *req); | |
2638 | void ptlrpc_abort_inflight(struct obd_import *imp); | |
2639 | void ptlrpc_cleanup_imp(struct obd_import *imp); | |
2640 | void ptlrpc_abort_set(struct ptlrpc_request_set *set); | |
2641 | ||
2642 | struct ptlrpc_request_set *ptlrpc_prep_set(void); | |
2643 | struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func, | |
2644 | void *arg); | |
2645 | int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, | |
2646 | set_interpreter_func fn, void *data); | |
2647 | int ptlrpc_set_next_timeout(struct ptlrpc_request_set *); | |
2648 | int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set); | |
2649 | int ptlrpc_set_wait(struct ptlrpc_request_set *); | |
2650 | int ptlrpc_expired_set(void *data); | |
2651 | void ptlrpc_interrupted_set(void *data); | |
2652 | void ptlrpc_mark_interrupted(struct ptlrpc_request *req); | |
2653 | void ptlrpc_set_destroy(struct ptlrpc_request_set *); | |
2654 | void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *); | |
2655 | void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, | |
2656 | struct ptlrpc_request *req); | |
2657 | ||
2658 | void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool); | |
2659 | void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq); | |
2660 | ||
2661 | struct ptlrpc_request_pool * | |
2662 | ptlrpc_init_rq_pool(int, int, | |
2663 | void (*populate_pool)(struct ptlrpc_request_pool *, int)); | |
2664 | ||
2665 | void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req); | |
2666 | struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp, | |
2667 | const struct req_format *format); | |
2668 | struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp, | |
2669 | struct ptlrpc_request_pool *, | |
2670 | const struct req_format *format); | |
2671 | void ptlrpc_request_free(struct ptlrpc_request *request); | |
2672 | int ptlrpc_request_pack(struct ptlrpc_request *request, | |
2673 | __u32 version, int opcode); | |
2674 | struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, | |
2675 | const struct req_format *format, | |
2676 | __u32 version, int opcode); | |
2677 | int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, | |
2678 | __u32 version, int opcode, char **bufs, | |
2679 | struct ptlrpc_cli_ctx *ctx); | |
2680 | struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version, | |
2681 | int opcode, int count, __u32 *lengths, | |
2682 | char **bufs); | |
2683 | struct ptlrpc_request *ptlrpc_prep_req_pool(struct obd_import *imp, | |
2684 | __u32 version, int opcode, | |
2685 | int count, __u32 *lengths, char **bufs, | |
2686 | struct ptlrpc_request_pool *pool); | |
2687 | void ptlrpc_req_finished(struct ptlrpc_request *request); | |
2688 | void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request); | |
2689 | struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req); | |
2690 | struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, | |
2691 | unsigned npages, unsigned max_brw, | |
2692 | unsigned type, unsigned portal); | |
2693 | void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin); | |
2694 | static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk) | |
2695 | { | |
2696 | __ptlrpc_free_bulk(bulk, 1); | |
2697 | } | |
2698 | static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk) | |
2699 | { | |
2700 | __ptlrpc_free_bulk(bulk, 0); | |
2701 | } | |
2702 | void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc, | |
2703 | struct page *page, int pageoffset, int len, int); | |
2704 | static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc, | |
2705 | struct page *page, int pageoffset, | |
2706 | int len) | |
2707 | { | |
2708 | __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1); | |
2709 | } | |
2710 | ||
2711 | static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc, | |
2712 | struct page *page, int pageoffset, | |
2713 | int len) | |
2714 | { | |
2715 | __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0); | |
2716 | } | |
2717 | ||
2718 | void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, | |
2719 | struct obd_import *imp); | |
2720 | __u64 ptlrpc_next_xid(void); | |
2721 | __u64 ptlrpc_sample_next_xid(void); | |
2722 | __u64 ptlrpc_req_xid(struct ptlrpc_request *request); | |
2723 | ||
2724 | /* Set of routines to run a function in ptlrpcd context */ | |
2725 | void *ptlrpcd_alloc_work(struct obd_import *imp, | |
2726 | int (*cb)(const struct lu_env *, void *), void *data); | |
2727 | void ptlrpcd_destroy_work(void *handler); | |
2728 | int ptlrpcd_queue_work(void *handler); | |
2729 | ||
2730 | /** @} */ | |
2731 | struct ptlrpc_service_buf_conf { | |
2732 | /* nbufs is buffers # to allocate when growing the pool */ | |
2733 | unsigned int bc_nbufs; | |
2734 | /* buffer size to post */ | |
2735 | unsigned int bc_buf_size; | |
2736 | /* portal to listed for requests on */ | |
2737 | unsigned int bc_req_portal; | |
2738 | /* portal of where to send replies to */ | |
2739 | unsigned int bc_rep_portal; | |
2740 | /* maximum request size to be accepted for this service */ | |
2741 | unsigned int bc_req_max_size; | |
2742 | /* maximum reply size this service can ever send */ | |
2743 | unsigned int bc_rep_max_size; | |
2744 | }; | |
2745 | ||
2746 | struct ptlrpc_service_thr_conf { | |
2747 | /* threadname should be 8 characters or less - 6 will be added on */ | |
2748 | char *tc_thr_name; | |
2749 | /* threads increasing factor for each CPU */ | |
2750 | unsigned int tc_thr_factor; | |
2751 | /* service threads # to start on each partition while initializing */ | |
2752 | unsigned int tc_nthrs_init; | |
2753 | /* | |
2754 | * low water of threads # upper-limit on each partition while running, | |
2755 | * service availability may be impacted if threads number is lower | |
2756 | * than this value. It can be ZERO if the service doesn't require | |
2757 | * CPU affinity or there is only one partition. | |
2758 | */ | |
2759 | unsigned int tc_nthrs_base; | |
2760 | /* "soft" limit for total threads number */ | |
2761 | unsigned int tc_nthrs_max; | |
2762 | /* user specified threads number, it will be validated due to | |
2763 | * other members of this structure. */ | |
2764 | unsigned int tc_nthrs_user; | |
2765 | /* set NUMA node affinity for service threads */ | |
2766 | unsigned int tc_cpu_affinity; | |
2767 | /* Tags for lu_context associated with service thread */ | |
2768 | __u32 tc_ctx_tags; | |
2769 | }; | |
2770 | ||
2771 | struct ptlrpc_service_cpt_conf { | |
2772 | struct cfs_cpt_table *cc_cptable; | |
2773 | /* string pattern to describe CPTs for a service */ | |
2774 | char *cc_pattern; | |
2775 | }; | |
2776 | ||
2777 | struct ptlrpc_service_conf { | |
2778 | /* service name */ | |
2779 | char *psc_name; | |
2780 | /* soft watchdog timeout multiplifier to print stuck service traces */ | |
2781 | unsigned int psc_watchdog_factor; | |
2782 | /* buffer information */ | |
2783 | struct ptlrpc_service_buf_conf psc_buf; | |
2784 | /* thread information */ | |
2785 | struct ptlrpc_service_thr_conf psc_thr; | |
2786 | /* CPU partition information */ | |
2787 | struct ptlrpc_service_cpt_conf psc_cpt; | |
2788 | /* function table */ | |
2789 | struct ptlrpc_service_ops psc_ops; | |
2790 | }; | |
2791 | ||
2792 | /* ptlrpc/service.c */ | |
2793 | /** | |
2794 | * Server-side services API. Register/unregister service, request state | |
2795 | * management, service thread management | |
2796 | * | |
2797 | * @{ | |
2798 | */ | |
2799 | void ptlrpc_save_lock(struct ptlrpc_request *req, | |
2800 | struct lustre_handle *lock, int mode, int no_ack); | |
2801 | void ptlrpc_commit_replies(struct obd_export *exp); | |
2802 | void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs); | |
2803 | void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs); | |
2804 | int ptlrpc_hpreq_handler(struct ptlrpc_request *req); | |
2805 | struct ptlrpc_service *ptlrpc_register_service( | |
2806 | struct ptlrpc_service_conf *conf, | |
2807 | struct proc_dir_entry *proc_entry); | |
2808 | void ptlrpc_stop_all_threads(struct ptlrpc_service *svc); | |
2809 | ||
2810 | int ptlrpc_start_threads(struct ptlrpc_service *svc); | |
2811 | int ptlrpc_unregister_service(struct ptlrpc_service *service); | |
2812 | int liblustre_check_services(void *arg); | |
2813 | void ptlrpc_daemonize(char *name); | |
2814 | int ptlrpc_service_health_check(struct ptlrpc_service *); | |
2815 | void ptlrpc_server_drop_request(struct ptlrpc_request *req); | |
2816 | void ptlrpc_request_change_export(struct ptlrpc_request *req, | |
2817 | struct obd_export *export); | |
2818 | ||
2819 | int ptlrpc_hr_init(void); | |
2820 | void ptlrpc_hr_fini(void); | |
2821 | ||
2822 | /** @} */ | |
2823 | ||
2824 | /* ptlrpc/import.c */ | |
2825 | /** | |
2826 | * Import API | |
2827 | * @{ | |
2828 | */ | |
2829 | int ptlrpc_connect_import(struct obd_import *imp); | |
2830 | int ptlrpc_init_import(struct obd_import *imp); | |
2831 | int ptlrpc_disconnect_import(struct obd_import *imp, int noclose); | |
2832 | int ptlrpc_import_recovery_state_machine(struct obd_import *imp); | |
2833 | void deuuidify(char *uuid, const char *prefix, char **uuid_start, | |
2834 | int *uuid_len); | |
2835 | ||
2836 | /* ptlrpc/pack_generic.c */ | |
2837 | int ptlrpc_reconnect_import(struct obd_import *imp); | |
2838 | /** @} */ | |
2839 | ||
2840 | /** | |
2841 | * ptlrpc msg buffer and swab interface | |
2842 | * | |
2843 | * @{ | |
2844 | */ | |
2845 | int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout, | |
2846 | int index); | |
2847 | void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout, | |
2848 | int index); | |
2849 | int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len); | |
2850 | int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len); | |
2851 | ||
2852 | int lustre_msg_check_version(struct lustre_msg *msg, __u32 version); | |
2853 | void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens, | |
2854 | char **bufs); | |
2855 | int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count, | |
2856 | __u32 *lens, char **bufs); | |
2857 | int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens, | |
2858 | char **bufs); | |
2859 | int lustre_pack_reply_v2(struct ptlrpc_request *req, int count, | |
2860 | __u32 *lens, char **bufs, int flags); | |
2861 | #define LPRFL_EARLY_REPLY 1 | |
2862 | int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens, | |
2863 | char **bufs, int flags); | |
2864 | int lustre_shrink_msg(struct lustre_msg *msg, int segment, | |
2865 | unsigned int newlen, int move_data); | |
2866 | void lustre_free_reply_state(struct ptlrpc_reply_state *rs); | |
2867 | int __lustre_unpack_msg(struct lustre_msg *m, int len); | |
2868 | int lustre_msg_hdr_size(__u32 magic, int count); | |
2869 | int lustre_msg_size(__u32 magic, int count, __u32 *lengths); | |
2870 | int lustre_msg_size_v2(int count, __u32 *lengths); | |
2871 | int lustre_packed_msg_size(struct lustre_msg *msg); | |
2872 | int lustre_msg_early_size(void); | |
2873 | void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size); | |
2874 | void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen); | |
2875 | int lustre_msg_buflen(struct lustre_msg *m, int n); | |
2876 | void lustre_msg_set_buflen(struct lustre_msg *m, int n, int len); | |
2877 | int lustre_msg_bufcount(struct lustre_msg *m); | |
2878 | char *lustre_msg_string(struct lustre_msg *m, int n, int max_len); | |
2879 | __u32 lustre_msghdr_get_flags(struct lustre_msg *msg); | |
2880 | void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags); | |
2881 | __u32 lustre_msg_get_flags(struct lustre_msg *msg); | |
2882 | void lustre_msg_add_flags(struct lustre_msg *msg, int flags); | |
2883 | void lustre_msg_set_flags(struct lustre_msg *msg, int flags); | |
2884 | void lustre_msg_clear_flags(struct lustre_msg *msg, int flags); | |
2885 | __u32 lustre_msg_get_op_flags(struct lustre_msg *msg); | |
2886 | void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags); | |
2887 | void lustre_msg_set_op_flags(struct lustre_msg *msg, int flags); | |
2888 | struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg); | |
2889 | __u32 lustre_msg_get_type(struct lustre_msg *msg); | |
2890 | __u32 lustre_msg_get_version(struct lustre_msg *msg); | |
2891 | void lustre_msg_add_version(struct lustre_msg *msg, int version); | |
2892 | __u32 lustre_msg_get_opc(struct lustre_msg *msg); | |
2893 | __u64 lustre_msg_get_last_xid(struct lustre_msg *msg); | |
2894 | __u64 lustre_msg_get_last_committed(struct lustre_msg *msg); | |
2895 | __u64 *lustre_msg_get_versions(struct lustre_msg *msg); | |
2896 | __u64 lustre_msg_get_transno(struct lustre_msg *msg); | |
2897 | __u64 lustre_msg_get_slv(struct lustre_msg *msg); | |
2898 | __u32 lustre_msg_get_limit(struct lustre_msg *msg); | |
2899 | void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv); | |
2900 | void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit); | |
2901 | int lustre_msg_get_status(struct lustre_msg *msg); | |
2902 | __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg); | |
2903 | int lustre_msg_is_v1(struct lustre_msg *msg); | |
2904 | __u32 lustre_msg_get_magic(struct lustre_msg *msg); | |
2905 | __u32 lustre_msg_get_timeout(struct lustre_msg *msg); | |
2906 | __u32 lustre_msg_get_service_time(struct lustre_msg *msg); | |
2907 | char *lustre_msg_get_jobid(struct lustre_msg *msg); | |
2908 | __u32 lustre_msg_get_cksum(struct lustre_msg *msg); | |
2909 | #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0) | |
2910 | __u32 lustre_msg_calc_cksum(struct lustre_msg *msg, int compat18); | |
2911 | #else | |
2912 | # warning "remove checksum compatibility support for b1_8" | |
2913 | __u32 lustre_msg_calc_cksum(struct lustre_msg *msg); | |
2914 | #endif | |
2915 | void lustre_msg_set_handle(struct lustre_msg *msg,struct lustre_handle *handle); | |
2916 | void lustre_msg_set_type(struct lustre_msg *msg, __u32 type); | |
2917 | void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc); | |
2918 | void lustre_msg_set_last_xid(struct lustre_msg *msg, __u64 last_xid); | |
2919 | void lustre_msg_set_last_committed(struct lustre_msg *msg,__u64 last_committed); | |
2920 | void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions); | |
2921 | void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno); | |
2922 | void lustre_msg_set_status(struct lustre_msg *msg, __u32 status); | |
2923 | void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt); | |
2924 | void ptlrpc_req_set_repsize(struct ptlrpc_request *req, int count, __u32 *sizes); | |
2925 | void ptlrpc_request_set_replen(struct ptlrpc_request *req); | |
2926 | void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout); | |
2927 | void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time); | |
2928 | void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid); | |
2929 | void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum); | |
2930 | ||
2931 | static inline void | |
2932 | lustre_shrink_reply(struct ptlrpc_request *req, int segment, | |
2933 | unsigned int newlen, int move_data) | |
2934 | { | |
2935 | LASSERT(req->rq_reply_state); | |
2936 | LASSERT(req->rq_repmsg); | |
2937 | req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment, | |
2938 | newlen, move_data); | |
2939 | } | |
2d58de78 LW |
2940 | |
2941 | #ifdef CONFIG_LUSTRE_TRANSLATE_ERRNOS | |
2942 | ||
2943 | static inline int ptlrpc_status_hton(int h) | |
2944 | { | |
2945 | /* | |
2946 | * Positive errnos must be network errnos, such as LUSTRE_EDEADLK, | |
2947 | * ELDLM_LOCK_ABORTED, etc. | |
2948 | */ | |
2949 | if (h < 0) | |
2950 | return -lustre_errno_hton(-h); | |
2951 | else | |
2952 | return h; | |
2953 | } | |
2954 | ||
2955 | static inline int ptlrpc_status_ntoh(int n) | |
2956 | { | |
2957 | /* | |
2958 | * See the comment in ptlrpc_status_hton(). | |
2959 | */ | |
2960 | if (n < 0) | |
2961 | return -lustre_errno_ntoh(-n); | |
2962 | else | |
2963 | return n; | |
2964 | } | |
2965 | ||
2966 | #else | |
2967 | ||
2968 | #define ptlrpc_status_hton(h) (h) | |
2969 | #define ptlrpc_status_ntoh(n) (n) | |
2970 | ||
2971 | #endif | |
d7e09d03 PT |
2972 | /** @} */ |
2973 | ||
2974 | /** Change request phase of \a req to \a new_phase */ | |
2975 | static inline void | |
2976 | ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase) | |
2977 | { | |
2978 | if (req->rq_phase == new_phase) | |
2979 | return; | |
2980 | ||
2981 | if (new_phase == RQ_PHASE_UNREGISTERING) { | |
2982 | req->rq_next_phase = req->rq_phase; | |
2983 | if (req->rq_import) | |
2984 | atomic_inc(&req->rq_import->imp_unregistering); | |
2985 | } | |
2986 | ||
2987 | if (req->rq_phase == RQ_PHASE_UNREGISTERING) { | |
2988 | if (req->rq_import) | |
2989 | atomic_dec(&req->rq_import->imp_unregistering); | |
2990 | } | |
2991 | ||
2992 | DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"", | |
2993 | ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase)); | |
2994 | ||
2995 | req->rq_phase = new_phase; | |
2996 | } | |
2997 | ||
2998 | /** | |
2999 | * Returns true if request \a req got early reply and hard deadline is not met | |
3000 | */ | |
3001 | static inline int | |
3002 | ptlrpc_client_early(struct ptlrpc_request *req) | |
3003 | { | |
3004 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
7264b8a5 | 3005 | req->rq_reply_deadline > get_seconds()) |
d7e09d03 PT |
3006 | return 0; |
3007 | return req->rq_early; | |
3008 | } | |
3009 | ||
3010 | /** | |
3011 | * Returns true if we got real reply from server for this request | |
3012 | */ | |
3013 | static inline int | |
3014 | ptlrpc_client_replied(struct ptlrpc_request *req) | |
3015 | { | |
3016 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
7264b8a5 | 3017 | req->rq_reply_deadline > get_seconds()) |
d7e09d03 PT |
3018 | return 0; |
3019 | return req->rq_replied; | |
3020 | } | |
3021 | ||
3022 | /** Returns true if request \a req is in process of receiving server reply */ | |
3023 | static inline int | |
3024 | ptlrpc_client_recv(struct ptlrpc_request *req) | |
3025 | { | |
3026 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
7264b8a5 | 3027 | req->rq_reply_deadline > get_seconds()) |
d7e09d03 PT |
3028 | return 1; |
3029 | return req->rq_receiving_reply; | |
3030 | } | |
3031 | ||
3032 | static inline int | |
3033 | ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req) | |
3034 | { | |
3035 | int rc; | |
3036 | ||
3037 | spin_lock(&req->rq_lock); | |
3038 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
7264b8a5 | 3039 | req->rq_reply_deadline > get_seconds()) { |
d7e09d03 PT |
3040 | spin_unlock(&req->rq_lock); |
3041 | return 1; | |
3042 | } | |
cf378ff7 AL |
3043 | rc = req->rq_receiving_reply; |
3044 | rc = rc || req->rq_req_unlink || req->rq_reply_unlink; | |
d7e09d03 PT |
3045 | spin_unlock(&req->rq_lock); |
3046 | return rc; | |
3047 | } | |
3048 | ||
3049 | static inline void | |
3050 | ptlrpc_client_wake_req(struct ptlrpc_request *req) | |
3051 | { | |
3052 | if (req->rq_set == NULL) | |
3053 | wake_up(&req->rq_reply_waitq); | |
3054 | else | |
3055 | wake_up(&req->rq_set->set_waitq); | |
3056 | } | |
3057 | ||
3058 | static inline void | |
3059 | ptlrpc_rs_addref(struct ptlrpc_reply_state *rs) | |
3060 | { | |
3061 | LASSERT(atomic_read(&rs->rs_refcount) > 0); | |
3062 | atomic_inc(&rs->rs_refcount); | |
3063 | } | |
3064 | ||
3065 | static inline void | |
3066 | ptlrpc_rs_decref(struct ptlrpc_reply_state *rs) | |
3067 | { | |
3068 | LASSERT(atomic_read(&rs->rs_refcount) > 0); | |
3069 | if (atomic_dec_and_test(&rs->rs_refcount)) | |
3070 | lustre_free_reply_state(rs); | |
3071 | } | |
3072 | ||
3073 | /* Should only be called once per req */ | |
3074 | static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req) | |
3075 | { | |
3076 | if (req->rq_reply_state == NULL) | |
3077 | return; /* shouldn't occur */ | |
3078 | ptlrpc_rs_decref(req->rq_reply_state); | |
3079 | req->rq_reply_state = NULL; | |
3080 | req->rq_repmsg = NULL; | |
3081 | } | |
3082 | ||
3083 | static inline __u32 lustre_request_magic(struct ptlrpc_request *req) | |
3084 | { | |
3085 | return lustre_msg_get_magic(req->rq_reqmsg); | |
3086 | } | |
3087 | ||
3088 | static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req) | |
3089 | { | |
3090 | switch (req->rq_reqmsg->lm_magic) { | |
3091 | case LUSTRE_MSG_MAGIC_V2: | |
3092 | return req->rq_reqmsg->lm_repsize; | |
3093 | default: | |
3094 | LASSERTF(0, "incorrect message magic: %08x\n", | |
3095 | req->rq_reqmsg->lm_magic); | |
3096 | return -EFAULT; | |
3097 | } | |
3098 | } | |
3099 | ||
3100 | static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req) | |
3101 | { | |
3102 | if (req->rq_delay_limit != 0 && | |
3103 | cfs_time_before(cfs_time_add(req->rq_queued_time, | |
3104 | cfs_time_seconds(req->rq_delay_limit)), | |
3105 | cfs_time_current())) { | |
3106 | return 1; | |
3107 | } | |
3108 | return 0; | |
3109 | } | |
3110 | ||
3111 | static inline int ptlrpc_no_resend(struct ptlrpc_request *req) | |
3112 | { | |
3113 | if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) { | |
3114 | spin_lock(&req->rq_lock); | |
3115 | req->rq_no_resend = 1; | |
3116 | spin_unlock(&req->rq_lock); | |
3117 | } | |
3118 | return req->rq_no_resend; | |
3119 | } | |
3120 | ||
3121 | static inline int | |
3122 | ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt) | |
3123 | { | |
3124 | int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate); | |
3125 | ||
3126 | return svcpt->scp_service->srv_watchdog_factor * | |
3127 | max_t(int, at, obd_timeout); | |
3128 | } | |
3129 | ||
3130 | static inline struct ptlrpc_service * | |
3131 | ptlrpc_req2svc(struct ptlrpc_request *req) | |
3132 | { | |
3133 | LASSERT(req->rq_rqbd != NULL); | |
3134 | return req->rq_rqbd->rqbd_svcpt->scp_service; | |
3135 | } | |
3136 | ||
3137 | /* ldlm/ldlm_lib.c */ | |
3138 | /** | |
3139 | * Target client logic | |
3140 | * @{ | |
3141 | */ | |
3142 | int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg); | |
3143 | int client_obd_cleanup(struct obd_device *obddev); | |
3144 | int client_connect_import(const struct lu_env *env, | |
3145 | struct obd_export **exp, struct obd_device *obd, | |
3146 | struct obd_uuid *cluuid, struct obd_connect_data *, | |
3147 | void *localdata); | |
3148 | int client_disconnect_export(struct obd_export *exp); | |
3149 | int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid, | |
3150 | int priority); | |
3151 | int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid); | |
3152 | int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, | |
3153 | struct obd_uuid *uuid); | |
3154 | int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid); | |
3155 | void client_destroy_import(struct obd_import *imp); | |
3156 | /** @} */ | |
3157 | ||
3158 | ||
3159 | /* ptlrpc/pinger.c */ | |
3160 | /** | |
3161 | * Pinger API (client side only) | |
3162 | * @{ | |
3163 | */ | |
3164 | enum timeout_event { | |
3165 | TIMEOUT_GRANT = 1 | |
3166 | }; | |
3167 | struct timeout_item; | |
3168 | typedef int (*timeout_cb_t)(struct timeout_item *, void *); | |
3169 | int ptlrpc_pinger_add_import(struct obd_import *imp); | |
3170 | int ptlrpc_pinger_del_import(struct obd_import *imp); | |
3171 | int ptlrpc_add_timeout_client(int time, enum timeout_event event, | |
3172 | timeout_cb_t cb, void *data, | |
3173 | struct list_head *obd_list); | |
3174 | int ptlrpc_del_timeout_client(struct list_head *obd_list, | |
3175 | enum timeout_event event); | |
3176 | struct ptlrpc_request * ptlrpc_prep_ping(struct obd_import *imp); | |
3177 | int ptlrpc_obd_ping(struct obd_device *obd); | |
d7e09d03 PT |
3178 | void ping_evictor_start(void); |
3179 | void ping_evictor_stop(void); | |
d7e09d03 PT |
3180 | void ptlrpc_pinger_ir_up(void); |
3181 | void ptlrpc_pinger_ir_down(void); | |
3182 | /** @} */ | |
3183 | int ptlrpc_pinger_suppress_pings(void); | |
3184 | ||
3185 | /* ptlrpc daemon bind policy */ | |
3186 | typedef enum { | |
3187 | /* all ptlrpcd threads are free mode */ | |
3188 | PDB_POLICY_NONE = 1, | |
3189 | /* all ptlrpcd threads are bound mode */ | |
3190 | PDB_POLICY_FULL = 2, | |
3191 | /* <free1 bound1> <free2 bound2> ... <freeN boundN> */ | |
3192 | PDB_POLICY_PAIR = 3, | |
3193 | /* <free1 bound1> <bound1 free2> ... <freeN boundN> <boundN free1>, | |
3194 | * means each ptlrpcd[X] has two partners: thread[X-1] and thread[X+1]. | |
3195 | * If kernel supports NUMA, pthrpcd threads are binded and | |
3196 | * grouped by NUMA node */ | |
3197 | PDB_POLICY_NEIGHBOR = 4, | |
3198 | } pdb_policy_t; | |
3199 | ||
3200 | /* ptlrpc daemon load policy | |
3201 | * It is caller's duty to specify how to push the async RPC into some ptlrpcd | |
3202 | * queue, but it is not enforced, affected by "ptlrpcd_bind_policy". If it is | |
3203 | * "PDB_POLICY_FULL", then the RPC will be processed by the selected ptlrpcd, | |
3204 | * Otherwise, the RPC may be processed by the selected ptlrpcd or its partner, | |
3205 | * depends on which is scheduled firstly, to accelerate the RPC processing. */ | |
3206 | typedef enum { | |
3207 | /* on the same CPU core as the caller */ | |
3208 | PDL_POLICY_SAME = 1, | |
3209 | /* within the same CPU partition, but not the same core as the caller */ | |
3210 | PDL_POLICY_LOCAL = 2, | |
3211 | /* round-robin on all CPU cores, but not the same core as the caller */ | |
3212 | PDL_POLICY_ROUND = 3, | |
3213 | /* the specified CPU core is preferred, but not enforced */ | |
3214 | PDL_POLICY_PREFERRED = 4, | |
3215 | } pdl_policy_t; | |
3216 | ||
3217 | /* ptlrpc/ptlrpcd.c */ | |
3218 | void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force); | |
3219 | void ptlrpcd_free(struct ptlrpcd_ctl *pc); | |
3220 | void ptlrpcd_wake(struct ptlrpc_request *req); | |
3221 | void ptlrpcd_add_req(struct ptlrpc_request *req, pdl_policy_t policy, int idx); | |
3222 | void ptlrpcd_add_rqset(struct ptlrpc_request_set *set); | |
3223 | int ptlrpcd_addref(void); | |
3224 | void ptlrpcd_decref(void); | |
3225 | ||
3226 | /* ptlrpc/lproc_ptlrpc.c */ | |
3227 | /** | |
3228 | * procfs output related functions | |
3229 | * @{ | |
3230 | */ | |
3231 | const char* ll_opcode2str(__u32 opcode); | |
3232 | #ifdef LPROCFS | |
3233 | void ptlrpc_lprocfs_register_obd(struct obd_device *obd); | |
3234 | void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd); | |
3235 | void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes); | |
3236 | #else | |
3237 | static inline void ptlrpc_lprocfs_register_obd(struct obd_device *obd) {} | |
3238 | static inline void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd) {} | |
3239 | static inline void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes) {} | |
3240 | #endif | |
3241 | /** @} */ | |
3242 | ||
d7e09d03 PT |
3243 | /* ptlrpc/llog_client.c */ |
3244 | extern struct llog_operations llog_client_ops; | |
3245 | ||
3246 | /** @} net */ | |
3247 | ||
3248 | #endif | |
3249 | /** @} PtlRPC */ |