]>
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 | * | |
1dc563a6 | 30 | * Copyright (c) 2010, 2015, Intel Corporation. |
d7e09d03 PT |
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 | ||
d7e09d03 PT |
49 | #ifndef _LUSTRE_NET_H |
50 | #define _LUSTRE_NET_H | |
51 | ||
52 | /** \defgroup net net | |
53 | * | |
54 | * @{ | |
55 | */ | |
56 | ||
9fdaf8c0 | 57 | #include "../../include/linux/libcfs/libcfs.h" |
bbf00c3d JS |
58 | #include "../../include/linux/lnet/nidstr.h" |
59 | #include "../../include/linux/lnet/api.h" | |
1accaadf GKH |
60 | #include "lustre/lustre_idl.h" |
61 | #include "lustre_ha.h" | |
62 | #include "lustre_sec.h" | |
63 | #include "lustre_import.h" | |
64 | #include "lprocfs_status.h" | |
65 | #include "lu_object.h" | |
66 | #include "lustre_req_layout.h" | |
d7e09d03 | 67 | |
1accaadf GKH |
68 | #include "obd_support.h" |
69 | #include "lustre_ver.h" | |
d7e09d03 PT |
70 | |
71 | /* MD flags we _always_ use */ | |
72 | #define PTLRPC_MD_OPTIONS 0 | |
73 | ||
74 | /** | |
75 | * Max # of bulk operations in one request. | |
76 | * In order for the client and server to properly negotiate the maximum | |
77 | * possible transfer size, PTLRPC_BULK_OPS_COUNT must be a power-of-two | |
78 | * value. The client is free to limit the actual RPC size for any bulk | |
c56e256d OD |
79 | * transfer via cl_max_pages_per_rpc to some non-power-of-two value. |
80 | */ | |
d7e09d03 PT |
81 | #define PTLRPC_BULK_OPS_BITS 2 |
82 | #define PTLRPC_BULK_OPS_COUNT (1U << PTLRPC_BULK_OPS_BITS) | |
83 | /** | |
84 | * PTLRPC_BULK_OPS_MASK is for the convenience of the client only, and | |
85 | * should not be used on the server at all. Otherwise, it imposes a | |
86 | * protocol limitation on the maximum RPC size that can be used by any | |
87 | * RPC sent to that server in the future. Instead, the server should | |
88 | * use the negotiated per-client ocd_brw_size to determine the bulk | |
c56e256d OD |
89 | * RPC count. |
90 | */ | |
d7e09d03 PT |
91 | #define PTLRPC_BULK_OPS_MASK (~((__u64)PTLRPC_BULK_OPS_COUNT - 1)) |
92 | ||
93 | /** | |
94 | * Define maxima for bulk I/O. | |
95 | * | |
96 | * A single PTLRPC BRW request is sent via up to PTLRPC_BULK_OPS_COUNT | |
97 | * of LNET_MTU sized RDMA transfers. Clients and servers negotiate the | |
98 | * currently supported maximum between peers at connect via ocd_brw_size. | |
99 | */ | |
100 | #define PTLRPC_MAX_BRW_BITS (LNET_MTU_BITS + PTLRPC_BULK_OPS_BITS) | |
101 | #define PTLRPC_MAX_BRW_SIZE (1 << PTLRPC_MAX_BRW_BITS) | |
09cbfeaf | 102 | #define PTLRPC_MAX_BRW_PAGES (PTLRPC_MAX_BRW_SIZE >> PAGE_SHIFT) |
d7e09d03 PT |
103 | |
104 | #define ONE_MB_BRW_SIZE (1 << LNET_MTU_BITS) | |
105 | #define MD_MAX_BRW_SIZE (1 << LNET_MTU_BITS) | |
09cbfeaf | 106 | #define MD_MAX_BRW_PAGES (MD_MAX_BRW_SIZE >> PAGE_SHIFT) |
d7e09d03 | 107 | #define DT_MAX_BRW_SIZE PTLRPC_MAX_BRW_SIZE |
09cbfeaf | 108 | #define DT_MAX_BRW_PAGES (DT_MAX_BRW_SIZE >> PAGE_SHIFT) |
d7e09d03 PT |
109 | #define OFD_MAX_BRW_SIZE (1 << LNET_MTU_BITS) |
110 | ||
111 | /* When PAGE_SIZE is a constant, we can check our arithmetic here with cpp! */ | |
112 | # if ((PTLRPC_MAX_BRW_PAGES & (PTLRPC_MAX_BRW_PAGES - 1)) != 0) | |
113 | # error "PTLRPC_MAX_BRW_PAGES isn't a power of two" | |
114 | # endif | |
ea1754a0 KS |
115 | # if (PTLRPC_MAX_BRW_SIZE != (PTLRPC_MAX_BRW_PAGES * PAGE_SIZE)) |
116 | # error "PTLRPC_MAX_BRW_SIZE isn't PTLRPC_MAX_BRW_PAGES * PAGE_SIZE" | |
d7e09d03 PT |
117 | # endif |
118 | # if (PTLRPC_MAX_BRW_SIZE > LNET_MTU * PTLRPC_BULK_OPS_COUNT) | |
119 | # error "PTLRPC_MAX_BRW_SIZE too big" | |
120 | # endif | |
121 | # if (PTLRPC_MAX_BRW_PAGES > LNET_MAX_IOV * PTLRPC_BULK_OPS_COUNT) | |
122 | # error "PTLRPC_MAX_BRW_PAGES too big" | |
123 | # endif | |
124 | ||
125 | #define PTLRPC_NTHRS_INIT 2 | |
126 | ||
127 | /** | |
128 | * Buffer Constants | |
129 | * | |
130 | * Constants determine how memory is used to buffer incoming service requests. | |
131 | * | |
132 | * ?_NBUFS # buffers to allocate when growing the pool | |
133 | * ?_BUFSIZE # bytes in a single request buffer | |
134 | * ?_MAXREQSIZE # maximum request service will receive | |
135 | * | |
136 | * When fewer than ?_NBUFS/2 buffers are posted for receive, another chunk | |
137 | * of ?_NBUFS is added to the pool. | |
138 | * | |
139 | * Messages larger than ?_MAXREQSIZE are dropped. Request buffers are | |
140 | * considered full when less than ?_MAXREQSIZE is left in them. | |
141 | */ | |
142 | /** | |
143 | * Thread Constants | |
144 | * | |
145 | * Constants determine how threads are created for ptlrpc service. | |
146 | * | |
147 | * ?_NTHRS_INIT # threads to create for each service partition on | |
148 | * initializing. If it's non-affinity service and | |
149 | * there is only one partition, it's the overall # | |
150 | * threads for the service while initializing. | |
151 | * ?_NTHRS_BASE # threads should be created at least for each | |
152 | * ptlrpc partition to keep the service healthy. | |
153 | * It's the low-water mark of threads upper-limit | |
154 | * for each partition. | |
155 | * ?_THR_FACTOR # threads can be added on threads upper-limit for | |
156 | * each CPU core. This factor is only for reference, | |
157 | * we might decrease value of factor if number of cores | |
158 | * per CPT is above a limit. | |
159 | * ?_NTHRS_MAX # overall threads can be created for a service, | |
160 | * it's a soft limit because if service is running | |
161 | * on machine with hundreds of cores and tens of | |
162 | * CPU partitions, we need to guarantee each partition | |
163 | * has ?_NTHRS_BASE threads, which means total threads | |
164 | * will be ?_NTHRS_BASE * number_of_cpts which can | |
165 | * exceed ?_NTHRS_MAX. | |
166 | * | |
167 | * Examples | |
168 | * | |
169 | * #define MDS_NTHRS_INIT 2 | |
170 | * #define MDS_NTHRS_BASE 64 | |
171 | * #define MDS_NTHRS_FACTOR 8 | |
172 | * #define MDS_NTHRS_MAX 1024 | |
173 | * | |
174 | * Example 1): | |
175 | * --------------------------------------------------------------------- | |
176 | * Server(A) has 16 cores, user configured it to 4 partitions so each | |
177 | * partition has 4 cores, then actual number of service threads on each | |
178 | * partition is: | |
179 | * MDS_NTHRS_BASE(64) + cores(4) * MDS_NTHRS_FACTOR(8) = 96 | |
180 | * | |
181 | * Total number of threads for the service is: | |
182 | * 96 * partitions(4) = 384 | |
183 | * | |
184 | * Example 2): | |
185 | * --------------------------------------------------------------------- | |
186 | * Server(B) has 32 cores, user configured it to 4 partitions so each | |
187 | * partition has 8 cores, then actual number of service threads on each | |
188 | * partition is: | |
189 | * MDS_NTHRS_BASE(64) + cores(8) * MDS_NTHRS_FACTOR(8) = 128 | |
190 | * | |
191 | * Total number of threads for the service is: | |
192 | * 128 * partitions(4) = 512 | |
193 | * | |
194 | * Example 3): | |
195 | * --------------------------------------------------------------------- | |
196 | * Server(B) has 96 cores, user configured it to 8 partitions so each | |
197 | * partition has 12 cores, then actual number of service threads on each | |
198 | * partition is: | |
199 | * MDS_NTHRS_BASE(64) + cores(12) * MDS_NTHRS_FACTOR(8) = 160 | |
200 | * | |
201 | * Total number of threads for the service is: | |
202 | * 160 * partitions(8) = 1280 | |
203 | * | |
204 | * However, it's above the soft limit MDS_NTHRS_MAX, so we choose this number | |
205 | * as upper limit of threads number for each partition: | |
206 | * MDS_NTHRS_MAX(1024) / partitions(8) = 128 | |
207 | * | |
208 | * Example 4): | |
209 | * --------------------------------------------------------------------- | |
210 | * Server(C) have a thousand of cores and user configured it to 32 partitions | |
211 | * MDS_NTHRS_BASE(64) * 32 = 2048 | |
212 | * | |
213 | * which is already above soft limit MDS_NTHRS_MAX(1024), but we still need | |
214 | * to guarantee that each partition has at least MDS_NTHRS_BASE(64) threads | |
215 | * to keep service healthy, so total number of threads will just be 2048. | |
216 | * | |
217 | * NB: we don't suggest to choose server with that many cores because backend | |
218 | * filesystem itself, buffer cache, or underlying network stack might | |
219 | * have some SMP scalability issues at that large scale. | |
220 | * | |
221 | * If user already has a fat machine with hundreds or thousands of cores, | |
222 | * there are two choices for configuration: | |
223 | * a) create CPU table from subset of all CPUs and run Lustre on | |
224 | * top of this subset | |
225 | * b) bind service threads on a few partitions, see modparameters of | |
226 | * MDS and OSS for details | |
227 | * | |
228 | * NB: these calculations (and examples below) are simplified to help | |
229 | * understanding, the real implementation is a little more complex, | |
230 | * please see ptlrpc_server_nthreads_check() for details. | |
231 | * | |
232 | */ | |
233 | ||
234 | /* | |
235 | * LDLM threads constants: | |
236 | * | |
237 | * Given 8 as factor and 24 as base threads number | |
238 | * | |
239 | * example 1) | |
240 | * On 4-core machine we will have 24 + 8 * 4 = 56 threads. | |
241 | * | |
242 | * example 2) | |
243 | * On 8-core machine with 2 partitions we will have 24 + 4 * 8 = 56 | |
244 | * threads for each partition and total threads number will be 112. | |
245 | * | |
246 | * example 3) | |
247 | * On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24) | |
248 | * threads for each partition to keep service healthy, so total threads | |
249 | * number should be 24 * 8 = 192. | |
250 | * | |
251 | * So with these constants, threads number will be at the similar level | |
252 | * of old versions, unless target machine has over a hundred cores | |
253 | */ | |
254 | #define LDLM_THR_FACTOR 8 | |
255 | #define LDLM_NTHRS_INIT PTLRPC_NTHRS_INIT | |
256 | #define LDLM_NTHRS_BASE 24 | |
257 | #define LDLM_NTHRS_MAX (num_online_cpus() == 1 ? 64 : 128) | |
258 | ||
259 | #define LDLM_BL_THREADS LDLM_NTHRS_AUTO_INIT | |
260 | #define LDLM_CLIENT_NBUFS 1 | |
261 | #define LDLM_SERVER_NBUFS 64 | |
262 | #define LDLM_BUFSIZE (8 * 1024) | |
263 | #define LDLM_MAXREQSIZE (5 * 1024) | |
264 | #define LDLM_MAXREPSIZE (1024) | |
265 | ||
e55c4476 JN |
266 | #define MDS_MAXREQSIZE (5 * 1024) /* >= 4736 */ |
267 | ||
d7e09d03 | 268 | #define OST_MAXREQSIZE (5 * 1024) |
d7e09d03 PT |
269 | |
270 | /* Macro to hide a typecast. */ | |
271 | #define ptlrpc_req_async_args(req) ((void *)&req->rq_async_args) | |
272 | ||
273 | /** | |
274 | * Structure to single define portal connection. | |
275 | */ | |
276 | struct ptlrpc_connection { | |
277 | /** linkage for connections hash table */ | |
278 | struct hlist_node c_hash; | |
279 | /** Our own lnet nid for this connection */ | |
280 | lnet_nid_t c_self; | |
281 | /** Remote side nid for this connection */ | |
282 | lnet_process_id_t c_peer; | |
283 | /** UUID of the other side */ | |
284 | struct obd_uuid c_remote_uuid; | |
285 | /** reference counter for this connection */ | |
286 | atomic_t c_refcount; | |
287 | }; | |
288 | ||
289 | /** Client definition for PortalRPC */ | |
290 | struct ptlrpc_client { | |
291 | /** What lnet portal does this client send messages to by default */ | |
292 | __u32 cli_request_portal; | |
293 | /** What portal do we expect replies on */ | |
294 | __u32 cli_reply_portal; | |
295 | /** Name of the client */ | |
296 | char *cli_name; | |
297 | }; | |
298 | ||
299 | /** state flags of requests */ | |
300 | /* XXX only ones left are those used by the bulk descs as well! */ | |
301 | #define PTL_RPC_FL_INTR (1 << 0) /* reply wait was interrupted by user */ | |
302 | #define PTL_RPC_FL_TIMEOUT (1 << 7) /* request timed out waiting for reply */ | |
303 | ||
304 | #define REQ_MAX_ACK_LOCKS 8 | |
305 | ||
306 | union ptlrpc_async_args { | |
307 | /** | |
308 | * Scratchpad for passing args to completion interpreter. Users | |
309 | * cast to the struct of their choosing, and CLASSERT that this is | |
c0a2472f | 310 | * big enough. For _tons_ of context, kmalloc a struct and store |
d7e09d03 PT |
311 | * a pointer to it here. The pointer_arg ensures this struct is at |
312 | * least big enough for that. | |
313 | */ | |
314 | void *pointer_arg[11]; | |
315 | __u64 space[7]; | |
316 | }; | |
317 | ||
318 | struct ptlrpc_request_set; | |
319 | typedef int (*set_interpreter_func)(struct ptlrpc_request_set *, void *, int); | |
320 | typedef int (*set_producer_func)(struct ptlrpc_request_set *, void *); | |
321 | ||
322 | /** | |
323 | * Definition of request set structure. | |
324 | * Request set is a list of requests (not necessary to the same target) that | |
325 | * once populated with RPCs could be sent in parallel. | |
326 | * There are two kinds of request sets. General purpose and with dedicated | |
327 | * serving thread. Example of the latter is ptlrpcd set. | |
328 | * For general purpose sets once request set started sending it is impossible | |
329 | * to add new requests to such set. | |
330 | * Provides a way to call "completion callbacks" when all requests in the set | |
331 | * returned. | |
332 | */ | |
333 | struct ptlrpc_request_set { | |
334 | atomic_t set_refcount; | |
335 | /** number of in queue requests */ | |
336 | atomic_t set_new_count; | |
337 | /** number of uncompleted requests */ | |
338 | atomic_t set_remaining; | |
339 | /** wait queue to wait on for request events */ | |
340 | wait_queue_head_t set_waitq; | |
341 | wait_queue_head_t *set_wakeup_ptr; | |
342 | /** List of requests in the set */ | |
343 | struct list_head set_requests; | |
344 | /** | |
345 | * List of completion callbacks to be called when the set is completed | |
346 | * This is only used if \a set_interpret is NULL. | |
347 | * Links struct ptlrpc_set_cbdata. | |
348 | */ | |
349 | struct list_head set_cblist; | |
350 | /** Completion callback, if only one. */ | |
351 | set_interpreter_func set_interpret; | |
352 | /** opaq argument passed to completion \a set_interpret callback. */ | |
353 | void *set_arg; | |
354 | /** | |
355 | * Lock for \a set_new_requests manipulations | |
356 | * locked so that any old caller can communicate requests to | |
357 | * the set holder who can then fold them into the lock-free set | |
358 | */ | |
359 | spinlock_t set_new_req_lock; | |
360 | /** List of new yet unsent requests. Only used with ptlrpcd now. */ | |
361 | struct list_head set_new_requests; | |
362 | ||
363 | /** rq_status of requests that have been freed already */ | |
364 | int set_rc; | |
365 | /** Additional fields used by the flow control extension */ | |
366 | /** Maximum number of RPCs in flight */ | |
367 | int set_max_inflight; | |
368 | /** Callback function used to generate RPCs */ | |
369 | set_producer_func set_producer; | |
370 | /** opaq argument passed to the producer callback */ | |
371 | void *set_producer_arg; | |
372 | }; | |
373 | ||
374 | /** | |
375 | * Description of a single ptrlrpc_set callback | |
376 | */ | |
377 | struct ptlrpc_set_cbdata { | |
378 | /** List linkage item */ | |
379 | struct list_head psc_item; | |
380 | /** Pointer to interpreting function */ | |
381 | set_interpreter_func psc_interpret; | |
382 | /** Opaq argument to pass to the callback */ | |
383 | void *psc_data; | |
384 | }; | |
385 | ||
386 | struct ptlrpc_bulk_desc; | |
387 | struct ptlrpc_service_part; | |
388 | struct ptlrpc_service; | |
389 | ||
390 | /** | |
391 | * ptlrpc callback & work item stuff | |
392 | */ | |
393 | struct ptlrpc_cb_id { | |
394 | void (*cbid_fn)(lnet_event_t *ev); /* specific callback fn */ | |
395 | void *cbid_arg; /* additional arg */ | |
396 | }; | |
397 | ||
398 | /** Maximum number of locks to fit into reply state */ | |
399 | #define RS_MAX_LOCKS 8 | |
400 | #define RS_DEBUG 0 | |
401 | ||
402 | /** | |
403 | * Structure to define reply state on the server | |
404 | * Reply state holds various reply message information. Also for "difficult" | |
405 | * replies (rep-ack case) we store the state after sending reply and wait | |
406 | * for the client to acknowledge the reception. In these cases locks could be | |
407 | * added to the state for replay/failover consistency guarantees. | |
408 | */ | |
409 | struct ptlrpc_reply_state { | |
410 | /** Callback description */ | |
411 | struct ptlrpc_cb_id rs_cb_id; | |
412 | /** Linkage for list of all reply states in a system */ | |
413 | struct list_head rs_list; | |
414 | /** Linkage for list of all reply states on same export */ | |
415 | struct list_head rs_exp_list; | |
416 | /** Linkage for list of all reply states for same obd */ | |
417 | struct list_head rs_obd_list; | |
418 | #if RS_DEBUG | |
419 | struct list_head rs_debug_list; | |
420 | #endif | |
421 | /** A spinlock to protect the reply state flags */ | |
422 | spinlock_t rs_lock; | |
423 | /** Reply state flags */ | |
c56e256d | 424 | unsigned long rs_difficult:1; /* ACK/commit stuff */ |
d7e09d03 | 425 | unsigned long rs_no_ack:1; /* no ACK, even for |
c56e256d OD |
426 | * difficult requests |
427 | */ | |
d7e09d03 PT |
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 | |
c56e256d OD |
434 | * and the rs was dispatched |
435 | */ | |
d7e09d03 PT |
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 */ | |
52ee0d20 | 470 | enum ldlm_mode rs_modes[RS_MAX_LOCKS]; |
d7e09d03 PT |
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 */ | |
aefd9d71 | 505 | int (*prp_populate)(struct ptlrpc_request_pool *, int); |
d7e09d03 PT |
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() | |
d7e09d03 PT |
648 | */ |
649 | int (*op_res_get) (struct ptlrpc_nrs_policy *policy, | |
650 | struct ptlrpc_nrs_request *nrq, | |
651 | const struct ptlrpc_nrs_resource *parent, | |
652 | struct ptlrpc_nrs_resource **resp, | |
653 | bool moving_req); | |
654 | /** | |
655 | * Called when releasing references taken for resources in the resource | |
656 | * hierarchy for the request; this operation is optional. | |
657 | * | |
658 | * \param[in,out] policy The policy the resource belongs to | |
659 | * \param[in] res The resource to be freed | |
660 | * | |
661 | * \see ptlrpc_nrs_req_finalize() | |
662 | * \see ptlrpc_nrs_hpreq_add_nolock() | |
d7e09d03 PT |
663 | */ |
664 | void (*op_res_put) (struct ptlrpc_nrs_policy *policy, | |
665 | const struct ptlrpc_nrs_resource *res); | |
666 | ||
667 | /** | |
668 | * Obtains a request for handling from the policy, and optionally | |
669 | * removes the request from the policy; this operation is mandatory. | |
670 | * | |
671 | * \param[in,out] policy The policy to poll | |
672 | * \param[in] peek When set, signifies that we just want to | |
673 | * examine the request, and not handle it, so the | |
674 | * request is not removed from the policy. | |
675 | * \param[in] force When set, it will force a policy to return a | |
676 | * request if it has one queued. | |
677 | * | |
678 | * \retval NULL No request available for handling | |
679 | * \retval valid-pointer The request polled for handling | |
680 | * | |
681 | * \see ptlrpc_nrs_req_get_nolock() | |
682 | */ | |
683 | struct ptlrpc_nrs_request * | |
684 | (*op_req_get) (struct ptlrpc_nrs_policy *policy, bool peek, | |
685 | bool force); | |
686 | /** | |
687 | * Called when attempting to add a request to a policy for later | |
688 | * handling; this operation is mandatory. | |
689 | * | |
690 | * \param[in,out] policy The policy on which to enqueue \a nrq | |
691 | * \param[in,out] nrq The request to enqueue | |
692 | * | |
693 | * \retval 0 success | |
694 | * \retval != 0 error | |
695 | * | |
696 | * \see ptlrpc_nrs_req_add_nolock() | |
697 | */ | |
698 | int (*op_req_enqueue) (struct ptlrpc_nrs_policy *policy, | |
699 | struct ptlrpc_nrs_request *nrq); | |
700 | /** | |
701 | * Removes a request from the policy's set of pending requests. Normally | |
702 | * called after a request has been polled successfully from the policy | |
703 | * for handling; this operation is mandatory. | |
704 | * | |
705 | * \param[in,out] policy The policy the request \a nrq belongs to | |
706 | * \param[in,out] nrq The request to dequeue | |
d7e09d03 PT |
707 | */ |
708 | void (*op_req_dequeue) (struct ptlrpc_nrs_policy *policy, | |
709 | struct ptlrpc_nrs_request *nrq); | |
710 | /** | |
711 | * Called after the request being carried out. Could be used for | |
712 | * job/resource control; this operation is optional. | |
713 | * | |
714 | * \param[in,out] policy The policy which is stopping to handle request | |
715 | * \a nrq | |
716 | * \param[in,out] nrq The request | |
717 | * | |
5e42bc9d | 718 | * \pre assert_spin_locked(&svcpt->scp_req_lock) |
d7e09d03 PT |
719 | * |
720 | * \see ptlrpc_nrs_req_stop_nolock() | |
721 | */ | |
722 | void (*op_req_stop) (struct ptlrpc_nrs_policy *policy, | |
723 | struct ptlrpc_nrs_request *nrq); | |
724 | /** | |
725 | * Registers the policy's lprocfs interface with a PTLRPC service. | |
726 | * | |
727 | * \param[in] svc The service | |
728 | * | |
729 | * \retval 0 success | |
730 | * \retval != 0 error | |
731 | */ | |
732 | int (*op_lprocfs_init) (struct ptlrpc_service *svc); | |
733 | /** | |
734 | * Unegisters the policy's lprocfs interface with a PTLRPC service. | |
735 | * | |
736 | * In cases of failed policy registration in | |
737 | * \e ptlrpc_nrs_policy_register(), this function may be called for a | |
738 | * service which has not registered the policy successfully, so | |
739 | * implementations of this method should make sure their operations are | |
740 | * safe in such cases. | |
741 | * | |
742 | * \param[in] svc The service | |
743 | */ | |
744 | void (*op_lprocfs_fini) (struct ptlrpc_service *svc); | |
745 | }; | |
746 | ||
747 | /** | |
748 | * Policy flags | |
749 | */ | |
750 | enum nrs_policy_flags { | |
751 | /** | |
752 | * Fallback policy, use this flag only on a single supported policy per | |
753 | * service. The flag cannot be used on policies that use | |
754 | * \e PTLRPC_NRS_FL_REG_EXTERN | |
755 | */ | |
756 | PTLRPC_NRS_FL_FALLBACK = (1 << 0), | |
757 | /** | |
758 | * Start policy immediately after registering. | |
759 | */ | |
760 | PTLRPC_NRS_FL_REG_START = (1 << 1), | |
761 | /** | |
762 | * This is a policy registering from a module different to the one NRS | |
763 | * core ships in (currently ptlrpc). | |
764 | */ | |
765 | PTLRPC_NRS_FL_REG_EXTERN = (1 << 2), | |
766 | }; | |
767 | ||
768 | /** | |
769 | * NRS queue type. | |
770 | * | |
771 | * Denotes whether an NRS instance is for handling normal or high-priority | |
772 | * RPCs, or whether an operation pertains to one or both of the NRS instances | |
773 | * in a service. | |
774 | */ | |
775 | enum ptlrpc_nrs_queue_type { | |
776 | PTLRPC_NRS_QUEUE_REG = (1 << 0), | |
777 | PTLRPC_NRS_QUEUE_HP = (1 << 1), | |
778 | PTLRPC_NRS_QUEUE_BOTH = (PTLRPC_NRS_QUEUE_REG | PTLRPC_NRS_QUEUE_HP) | |
779 | }; | |
780 | ||
781 | /** | |
782 | * NRS head | |
783 | * | |
784 | * A PTLRPC service has at least one NRS head instance for handling normal | |
785 | * priority RPCs, and may optionally have a second NRS head instance for | |
786 | * handling high-priority RPCs. Each NRS head maintains a list of available | |
787 | * policies, of which one and only one policy is acting as the fallback policy, | |
788 | * and optionally a different policy may be acting as the primary policy. For | |
789 | * all RPCs handled by this NRS head instance, NRS core will first attempt to | |
790 | * enqueue the RPC using the primary policy (if any). The fallback policy is | |
791 | * used in the following cases: | |
792 | * - when there was no primary policy in the | |
793 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state at the time the request | |
794 | * was initialized. | |
795 | * - when the primary policy that was at the | |
796 | * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the | |
797 | * RPC was initialized, denoted it did not wish, or for some other reason was | |
798 | * not able to handle the request, by returning a non-valid NRS resource | |
799 | * reference. | |
800 | * - when the primary policy that was at the | |
801 | * ptlrpc_nrs_pol_state::PTLRPC_NRS_POL_STATE_STARTED state at the time the | |
802 | * RPC was initialized, fails later during the request enqueueing stage. | |
803 | * | |
804 | * \see nrs_resource_get_safe() | |
805 | * \see nrs_request_enqueue() | |
806 | */ | |
807 | struct ptlrpc_nrs { | |
808 | spinlock_t nrs_lock; | |
809 | /** XXX Possibly replace svcpt->scp_req_lock with another lock here. */ | |
810 | /** | |
811 | * List of registered policies | |
812 | */ | |
813 | struct list_head nrs_policy_list; | |
814 | /** | |
815 | * List of policies with queued requests. Policies that have any | |
816 | * outstanding requests are queued here, and this list is queried | |
817 | * in a round-robin manner from NRS core when obtaining a request | |
818 | * for handling. This ensures that requests from policies that at some | |
819 | * point transition away from the | |
820 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED state are drained. | |
821 | */ | |
822 | struct list_head nrs_policy_queued; | |
823 | /** | |
824 | * Service partition for this NRS head | |
825 | */ | |
826 | struct ptlrpc_service_part *nrs_svcpt; | |
827 | /** | |
828 | * Primary policy, which is the preferred policy for handling RPCs | |
829 | */ | |
830 | struct ptlrpc_nrs_policy *nrs_policy_primary; | |
831 | /** | |
832 | * Fallback policy, which is the backup policy for handling RPCs | |
833 | */ | |
834 | struct ptlrpc_nrs_policy *nrs_policy_fallback; | |
835 | /** | |
836 | * This NRS head handles either HP or regular requests | |
837 | */ | |
838 | enum ptlrpc_nrs_queue_type nrs_queue_type; | |
839 | /** | |
840 | * # queued requests from all policies in this NRS head | |
841 | */ | |
842 | unsigned long nrs_req_queued; | |
843 | /** | |
844 | * # scheduled requests from all policies in this NRS head | |
845 | */ | |
846 | unsigned long nrs_req_started; | |
847 | /** | |
848 | * # policies on this NRS | |
849 | */ | |
850 | unsigned nrs_num_pols; | |
851 | /** | |
852 | * This NRS head is in progress of starting a policy | |
853 | */ | |
854 | unsigned nrs_policy_starting:1; | |
855 | /** | |
856 | * In progress of shutting down the whole NRS head; used during | |
857 | * unregistration | |
858 | */ | |
859 | unsigned nrs_stopping:1; | |
860 | }; | |
861 | ||
862 | #define NRS_POL_NAME_MAX 16 | |
863 | ||
864 | struct ptlrpc_nrs_pol_desc; | |
865 | ||
866 | /** | |
867 | * Service compatibility predicate; this determines whether a policy is adequate | |
868 | * for handling RPCs of a particular PTLRPC service. | |
869 | * | |
870 | * XXX:This should give the same result during policy registration and | |
871 | * unregistration, and for all partitions of a service; so the result should not | |
872 | * depend on temporal service or other properties, that may influence the | |
873 | * result. | |
874 | */ | |
875 | typedef bool (*nrs_pol_desc_compat_t) (const struct ptlrpc_service *svc, | |
876 | const struct ptlrpc_nrs_pol_desc *desc); | |
877 | ||
878 | struct ptlrpc_nrs_pol_conf { | |
879 | /** | |
880 | * Human-readable policy name | |
881 | */ | |
882 | char nc_name[NRS_POL_NAME_MAX]; | |
883 | /** | |
884 | * NRS operations for this policy | |
885 | */ | |
886 | const struct ptlrpc_nrs_pol_ops *nc_ops; | |
887 | /** | |
888 | * Service compatibility predicate | |
889 | */ | |
890 | nrs_pol_desc_compat_t nc_compat; | |
891 | /** | |
892 | * Set for policies that support a single ptlrpc service, i.e. ones that | |
893 | * have \a pd_compat set to nrs_policy_compat_one(). The variable value | |
894 | * depicts the name of the single service that such policies are | |
895 | * compatible with. | |
896 | */ | |
897 | const char *nc_compat_svc_name; | |
898 | /** | |
899 | * Owner module for this policy descriptor; policies registering from a | |
900 | * different module to the one the NRS framework is held within | |
901 | * (currently ptlrpc), should set this field to THIS_MODULE. | |
902 | */ | |
c34d9cd8 | 903 | struct module *nc_owner; |
d7e09d03 | 904 | /** |
bd9070cb | 905 | * Policy registration flags; a bitmask of \e nrs_policy_flags |
d7e09d03 PT |
906 | */ |
907 | unsigned nc_flags; | |
908 | }; | |
909 | ||
910 | /** | |
911 | * NRS policy registering descriptor | |
912 | * | |
913 | * Is used to hold a description of a policy that can be passed to NRS core in | |
914 | * order to register the policy with NRS heads in different PTLRPC services. | |
915 | */ | |
916 | struct ptlrpc_nrs_pol_desc { | |
917 | /** | |
918 | * Human-readable policy name | |
919 | */ | |
920 | char pd_name[NRS_POL_NAME_MAX]; | |
921 | /** | |
922 | * Link into nrs_core::nrs_policies | |
923 | */ | |
924 | struct list_head pd_list; | |
925 | /** | |
926 | * NRS operations for this policy | |
927 | */ | |
928 | const struct ptlrpc_nrs_pol_ops *pd_ops; | |
929 | /** | |
930 | * Service compatibility predicate | |
931 | */ | |
932 | nrs_pol_desc_compat_t pd_compat; | |
933 | /** | |
934 | * Set for policies that are compatible with only one PTLRPC service. | |
935 | * | |
936 | * \see ptlrpc_nrs_pol_conf::nc_compat_svc_name | |
937 | */ | |
938 | const char *pd_compat_svc_name; | |
939 | /** | |
940 | * Owner module for this policy descriptor. | |
941 | * | |
942 | * We need to hold a reference to the module whenever we might make use | |
943 | * of any of the module's contents, i.e. | |
944 | * - If one or more instances of the policy are at a state where they | |
945 | * might be handling a request, i.e. | |
946 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or | |
947 | * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING as we will have to | |
948 | * call into the policy's ptlrpc_nrs_pol_ops() handlers. A reference | |
949 | * is taken on the module when | |
950 | * \e ptlrpc_nrs_pol_desc::pd_refs becomes 1, and released when it | |
951 | * becomes 0, so that we hold only one reference to the module maximum | |
952 | * at any time. | |
953 | * | |
954 | * We do not need to hold a reference to the module, even though we | |
955 | * might use code and data from the module, in the following cases: | |
956 | * - During external policy registration, because this should happen in | |
957 | * the module's init() function, in which case the module is safe from | |
958 | * removal because a reference is being held on the module by the | |
959 | * kernel, and iirc kmod (and I guess module-init-tools also) will | |
960 | * serialize any racing processes properly anyway. | |
961 | * - During external policy unregistration, because this should happen | |
962 | * in a module's exit() function, and any attempts to start a policy | |
963 | * instance would need to take a reference on the module, and this is | |
964 | * not possible once we have reached the point where the exit() | |
965 | * handler is called. | |
966 | * - During service registration and unregistration, as service setup | |
967 | * and cleanup, and policy registration, unregistration and policy | |
968 | * instance starting, are serialized by \e nrs_core::nrs_mutex, so | |
969 | * as long as users adhere to the convention of registering policies | |
970 | * in init() and unregistering them in module exit() functions, there | |
971 | * should not be a race between these operations. | |
972 | * - During any policy-specific lprocfs operations, because a reference | |
973 | * is held by the kernel on a proc entry that has been entered by a | |
974 | * syscall, so as long as proc entries are removed during unregistration time, | |
975 | * then unregistration and lprocfs operations will be properly | |
976 | * serialized. | |
977 | */ | |
c34d9cd8 | 978 | struct module *pd_owner; |
d7e09d03 PT |
979 | /** |
980 | * Bitmask of \e nrs_policy_flags | |
981 | */ | |
982 | unsigned pd_flags; | |
983 | /** | |
984 | * # of references on this descriptor | |
985 | */ | |
986 | atomic_t pd_refs; | |
987 | }; | |
988 | ||
989 | /** | |
990 | * NRS policy state | |
991 | * | |
992 | * Policies transition from one state to the other during their lifetime | |
993 | */ | |
994 | enum ptlrpc_nrs_pol_state { | |
995 | /** | |
996 | * Not a valid policy state. | |
997 | */ | |
998 | NRS_POL_STATE_INVALID, | |
999 | /** | |
1000 | * Policies are at this state either at the start of their life, or | |
1001 | * transition here when the user selects a different policy to act | |
1002 | * as the primary one. | |
1003 | */ | |
1004 | NRS_POL_STATE_STOPPED, | |
1005 | /** | |
1006 | * Policy is progress of stopping | |
1007 | */ | |
1008 | NRS_POL_STATE_STOPPING, | |
1009 | /** | |
1010 | * Policy is in progress of starting | |
1011 | */ | |
1012 | NRS_POL_STATE_STARTING, | |
1013 | /** | |
1014 | * A policy is in this state in two cases: | |
1015 | * - it is the fallback policy, which is always in this state. | |
1016 | * - it has been activated by the user; i.e. it is the primary policy, | |
1017 | */ | |
1018 | NRS_POL_STATE_STARTED, | |
1019 | }; | |
1020 | ||
1021 | /** | |
1022 | * NRS policy information | |
1023 | * | |
1024 | * Used for obtaining information for the status of a policy via lprocfs | |
1025 | */ | |
1026 | struct ptlrpc_nrs_pol_info { | |
1027 | /** | |
1028 | * Policy name | |
1029 | */ | |
1030 | char pi_name[NRS_POL_NAME_MAX]; | |
1031 | /** | |
1032 | * Current policy state | |
1033 | */ | |
1034 | enum ptlrpc_nrs_pol_state pi_state; | |
1035 | /** | |
1036 | * # RPCs enqueued for later dispatching by the policy | |
1037 | */ | |
1038 | long pi_req_queued; | |
1039 | /** | |
1040 | * # RPCs started for dispatch by the policy | |
1041 | */ | |
1042 | long pi_req_started; | |
1043 | /** | |
1044 | * Is this a fallback policy? | |
1045 | */ | |
1046 | unsigned pi_fallback:1; | |
1047 | }; | |
1048 | ||
1049 | /** | |
1050 | * NRS policy | |
1051 | * | |
1052 | * There is one instance of this for each policy in each NRS head of each | |
1053 | * PTLRPC service partition. | |
1054 | */ | |
1055 | struct ptlrpc_nrs_policy { | |
1056 | /** | |
1057 | * Linkage into the NRS head's list of policies, | |
1058 | * ptlrpc_nrs:nrs_policy_list | |
1059 | */ | |
1060 | struct list_head pol_list; | |
1061 | /** | |
1062 | * Linkage into the NRS head's list of policies with enqueued | |
1063 | * requests ptlrpc_nrs:nrs_policy_queued | |
1064 | */ | |
1065 | struct list_head pol_list_queued; | |
1066 | /** | |
1067 | * Current state of this policy | |
1068 | */ | |
1069 | enum ptlrpc_nrs_pol_state pol_state; | |
1070 | /** | |
1071 | * Bitmask of nrs_policy_flags | |
1072 | */ | |
1073 | unsigned pol_flags; | |
1074 | /** | |
1075 | * # RPCs enqueued for later dispatching by the policy | |
1076 | */ | |
1077 | long pol_req_queued; | |
1078 | /** | |
1079 | * # RPCs started for dispatch by the policy | |
1080 | */ | |
1081 | long pol_req_started; | |
1082 | /** | |
1083 | * Usage Reference count taken on the policy instance | |
1084 | */ | |
1085 | long pol_ref; | |
1086 | /** | |
1087 | * The NRS head this policy has been created at | |
1088 | */ | |
1089 | struct ptlrpc_nrs *pol_nrs; | |
1090 | /** | |
1091 | * Private policy data; varies by policy type | |
1092 | */ | |
1093 | void *pol_private; | |
1094 | /** | |
1095 | * Policy descriptor for this policy instance. | |
1096 | */ | |
1097 | struct ptlrpc_nrs_pol_desc *pol_desc; | |
1098 | }; | |
1099 | ||
1100 | /** | |
1101 | * NRS resource | |
1102 | * | |
1103 | * Resources are embedded into two types of NRS entities: | |
1104 | * - Inside NRS policies, in the policy's private data in | |
1105 | * ptlrpc_nrs_policy::pol_private | |
1106 | * - In objects that act as prime-level scheduling entities in different NRS | |
1107 | * policies; e.g. on a policy that performs round robin or similar order | |
1108 | * scheduling across client NIDs, there would be one NRS resource per unique | |
1109 | * client NID. On a policy which performs round robin scheduling across | |
1110 | * backend filesystem objects, there would be one resource associated with | |
1111 | * each of the backend filesystem objects partaking in the scheduling | |
1112 | * performed by the policy. | |
1113 | * | |
1114 | * NRS resources share a parent-child relationship, in which resources embedded | |
1115 | * in policy instances are the parent entities, with all scheduling entities | |
1116 | * a policy schedules across being the children, thus forming a simple resource | |
1117 | * hierarchy. This hierarchy may be extended with one or more levels in the | |
1118 | * future if the ability to have more than one primary policy is added. | |
1119 | * | |
1120 | * Upon request initialization, references to the then active NRS policies are | |
1121 | * taken and used to later handle the dispatching of the request with one of | |
1122 | * these policies. | |
1123 | * | |
1124 | * \see nrs_resource_get_safe() | |
1125 | * \see ptlrpc_nrs_req_add() | |
1126 | */ | |
1127 | struct ptlrpc_nrs_resource { | |
1128 | /** | |
1129 | * This NRS resource's parent; is NULL for resources embedded in NRS | |
1130 | * policy instances; i.e. those are top-level ones. | |
1131 | */ | |
1132 | struct ptlrpc_nrs_resource *res_parent; | |
1133 | /** | |
1134 | * The policy associated with this resource. | |
1135 | */ | |
1136 | struct ptlrpc_nrs_policy *res_policy; | |
1137 | }; | |
1138 | ||
1139 | enum { | |
1140 | NRS_RES_FALLBACK, | |
1141 | NRS_RES_PRIMARY, | |
1142 | NRS_RES_MAX | |
1143 | }; | |
1144 | ||
1145 | /* \name fifo | |
1146 | * | |
1147 | * FIFO policy | |
1148 | * | |
1149 | * This policy is a logical wrapper around previous, non-NRS functionality. | |
1150 | * It dispatches RPCs in the same order as they arrive from the network. This | |
1151 | * policy is currently used as the fallback policy, and the only enabled policy | |
1152 | * on all NRS heads of all PTLRPC service partitions. | |
1153 | * @{ | |
1154 | */ | |
1155 | ||
1156 | /** | |
1157 | * Private data structure for the FIFO policy | |
1158 | */ | |
1159 | struct nrs_fifo_head { | |
1160 | /** | |
1161 | * Resource object for policy instance. | |
1162 | */ | |
1163 | struct ptlrpc_nrs_resource fh_res; | |
1164 | /** | |
1165 | * List of queued requests. | |
1166 | */ | |
1167 | struct list_head fh_list; | |
1168 | /** | |
1169 | * For debugging purposes. | |
1170 | */ | |
1171 | __u64 fh_sequence; | |
1172 | }; | |
1173 | ||
1174 | struct nrs_fifo_req { | |
1175 | struct list_head fr_list; | |
1176 | __u64 fr_sequence; | |
1177 | }; | |
1178 | ||
1179 | /** @} fifo */ | |
1180 | ||
d7e09d03 PT |
1181 | /** |
1182 | * NRS request | |
1183 | * | |
1184 | * Instances of this object exist embedded within ptlrpc_request; the main | |
1185 | * purpose of this object is to hold references to the request's resources | |
1186 | * for the lifetime of the request, and to hold properties that policies use | |
1187 | * use for determining the request's scheduling priority. | |
c56e256d | 1188 | */ |
d7e09d03 PT |
1189 | struct ptlrpc_nrs_request { |
1190 | /** | |
1191 | * The request's resource hierarchy. | |
1192 | */ | |
1193 | struct ptlrpc_nrs_resource *nr_res_ptrs[NRS_RES_MAX]; | |
1194 | /** | |
1195 | * Index into ptlrpc_nrs_request::nr_res_ptrs of the resource of the | |
1196 | * policy that was used to enqueue the request. | |
1197 | * | |
1198 | * \see nrs_request_enqueue() | |
1199 | */ | |
1200 | unsigned nr_res_idx; | |
1201 | unsigned nr_initialized:1; | |
1202 | unsigned nr_enqueued:1; | |
1203 | unsigned nr_started:1; | |
1204 | unsigned nr_finalized:1; | |
d7e09d03 PT |
1205 | |
1206 | /** | |
1207 | * Policy-specific fields, used for determining a request's scheduling | |
1208 | * priority, and other supporting functionality. | |
1209 | */ | |
1210 | union { | |
1211 | /** | |
1212 | * Fields for the FIFO policy | |
1213 | */ | |
1214 | struct nrs_fifo_req fifo; | |
d7e09d03 PT |
1215 | } nr_u; |
1216 | /** | |
1217 | * Externally-registering policies may want to use this to allocate | |
1218 | * their own request properties. | |
1219 | */ | |
1220 | void *ext; | |
1221 | }; | |
1222 | ||
1223 | /** @} nrs */ | |
1224 | ||
1225 | /** | |
1226 | * Basic request prioritization operations structure. | |
1227 | * The whole idea is centered around locks and RPCs that might affect locks. | |
1228 | * When a lock is contended we try to give priority to RPCs that might lead | |
1229 | * to fastest release of that lock. | |
1230 | * Currently only implemented for OSTs only in a way that makes all | |
1231 | * IO and truncate RPCs that are coming from a locked region where a lock is | |
1232 | * contended a priority over other requests. | |
1233 | */ | |
1234 | struct ptlrpc_hpreq_ops { | |
1235 | /** | |
1236 | * Check if the lock handle of the given lock is the same as | |
1237 | * taken from the request. | |
1238 | */ | |
1239 | int (*hpreq_lock_match)(struct ptlrpc_request *, struct ldlm_lock *); | |
1240 | /** | |
1241 | * Check if the request is a high priority one. | |
1242 | */ | |
1243 | int (*hpreq_check)(struct ptlrpc_request *); | |
1244 | /** | |
1245 | * Called after the request has been handled. | |
1246 | */ | |
1247 | void (*hpreq_fini)(struct ptlrpc_request *); | |
1248 | }; | |
1249 | ||
1250 | /** | |
1251 | * Represents remote procedure call. | |
1252 | * | |
1253 | * This is a staple structure used by everybody wanting to send a request | |
1254 | * in Lustre. | |
1255 | */ | |
1256 | struct ptlrpc_request { | |
1257 | /* Request type: one of PTL_RPC_MSG_* */ | |
1258 | int rq_type; | |
1259 | /** Result of request processing */ | |
1260 | int rq_status; | |
1261 | /** | |
1262 | * Linkage item through which this request is included into | |
1263 | * sending/delayed lists on client and into rqbd list on server | |
1264 | */ | |
1265 | struct list_head rq_list; | |
1266 | /** | |
1267 | * Server side list of incoming unserved requests sorted by arrival | |
1268 | * time. Traversed from time to time to notice about to expire | |
1269 | * requests and sent back "early replies" to clients to let them | |
1270 | * know server is alive and well, just very busy to service their | |
1271 | * requests in time | |
1272 | */ | |
1273 | struct list_head rq_timed_list; | |
17891183 | 1274 | /** server-side history, used for debugging purposes. */ |
d7e09d03 PT |
1275 | struct list_head rq_history_list; |
1276 | /** server-side per-export list */ | |
1277 | struct list_head rq_exp_list; | |
1278 | /** server-side hp handlers */ | |
1279 | struct ptlrpc_hpreq_ops *rq_ops; | |
1280 | ||
1281 | /** initial thread servicing this request */ | |
1282 | struct ptlrpc_thread *rq_svc_thread; | |
1283 | ||
1284 | /** history sequence # */ | |
1285 | __u64 rq_history_seq; | |
1286 | /** \addtogroup nrs | |
1287 | * @{ | |
1288 | */ | |
1289 | /** stub for NRS request */ | |
1290 | struct ptlrpc_nrs_request rq_nrq; | |
1291 | /** @} nrs */ | |
1292 | /** the index of service's srv_at_array into which request is linked */ | |
c8ff1810 | 1293 | u32 rq_at_index; |
d7e09d03 PT |
1294 | /** Lock to protect request flags and some other important bits, like |
1295 | * rq_list | |
1296 | */ | |
1297 | spinlock_t rq_lock; | |
1298 | /** client-side flags are serialized by rq_lock */ | |
1299 | unsigned int rq_intr:1, rq_replied:1, rq_err:1, | |
1300 | rq_timedout:1, rq_resend:1, rq_restart:1, | |
1301 | /** | |
1302 | * when ->rq_replay is set, request is kept by the client even | |
1303 | * after server commits corresponding transaction. This is | |
1304 | * used for operations that require sequence of multiple | |
1305 | * requests to be replayed. The only example currently is file | |
1306 | * open/close. When last request in such a sequence is | |
1307 | * committed, ->rq_replay is cleared on all requests in the | |
1308 | * sequence. | |
1309 | */ | |
1310 | rq_replay:1, | |
1311 | rq_no_resend:1, rq_waiting:1, rq_receiving_reply:1, | |
1312 | rq_no_delay:1, rq_net_err:1, rq_wait_ctx:1, | |
cf378ff7 AL |
1313 | rq_early:1, |
1314 | rq_req_unlink:1, rq_reply_unlink:1, | |
d7e09d03 PT |
1315 | rq_memalloc:1, /* req originated from "kswapd" */ |
1316 | /* server-side flags */ | |
1317 | rq_packed_final:1, /* packed final reply */ | |
1318 | rq_hp:1, /* high priority RPC */ | |
1319 | rq_at_linked:1, /* link into service's srv_at_array */ | |
1320 | rq_reply_truncate:1, | |
1321 | rq_committed:1, | |
1322 | /* whether the "rq_set" is a valid one */ | |
1323 | rq_invalid_rqset:1, | |
1324 | rq_generation_set:1, | |
1325 | /* do not resend request on -EINPROGRESS */ | |
1326 | rq_no_retry_einprogress:1, | |
1327 | /* allow the req to be sent if the import is in recovery | |
c56e256d OD |
1328 | * status |
1329 | */ | |
c52f69c5 | 1330 | rq_allow_replay:1; |
d7e09d03 PT |
1331 | |
1332 | unsigned int rq_nr_resend; | |
1333 | ||
1334 | enum rq_phase rq_phase; /* one of RQ_PHASE_* */ | |
1335 | enum rq_phase rq_next_phase; /* one of RQ_PHASE_* to be used next */ | |
c56e256d OD |
1336 | atomic_t rq_refcount; /* client-side refcount for SENT race, |
1337 | * server-side refcount for multiple replies | |
1338 | */ | |
d7e09d03 PT |
1339 | |
1340 | /** Portal to which this request would be sent */ | |
1341 | short rq_request_portal; /* XXX FIXME bug 249 */ | |
1342 | /** Portal where to wait for reply and where reply would be sent */ | |
1343 | short rq_reply_portal; /* XXX FIXME bug 249 */ | |
1344 | ||
1345 | /** | |
1346 | * client-side: | |
1347 | * !rq_truncate : # reply bytes actually received, | |
1348 | * rq_truncate : required repbuf_len for resend | |
1349 | */ | |
1350 | int rq_nob_received; | |
1351 | /** Request length */ | |
1352 | int rq_reqlen; | |
1353 | /** Reply length */ | |
1354 | int rq_replen; | |
1355 | /** Request message - what client sent */ | |
1356 | struct lustre_msg *rq_reqmsg; | |
1357 | /** Reply message - server response */ | |
1358 | struct lustre_msg *rq_repmsg; | |
1359 | /** Transaction number */ | |
1360 | __u64 rq_transno; | |
1361 | /** xid */ | |
1362 | __u64 rq_xid; | |
1363 | /** | |
17891183 | 1364 | * List item to for replay list. Not yet committed requests get linked |
d7e09d03 PT |
1365 | * there. |
1366 | * Also see \a rq_replay comment above. | |
1367 | */ | |
1368 | struct list_head rq_replay_list; | |
1369 | ||
1370 | /** | |
1371 | * security and encryption data | |
c56e256d OD |
1372 | * @{ |
1373 | */ | |
d7e09d03 PT |
1374 | struct ptlrpc_cli_ctx *rq_cli_ctx; /**< client's half ctx */ |
1375 | struct ptlrpc_svc_ctx *rq_svc_ctx; /**< server's half ctx */ | |
1376 | struct list_head rq_ctx_chain; /**< link to waited ctx */ | |
1377 | ||
1378 | struct sptlrpc_flavor rq_flvr; /**< for client & server */ | |
1379 | enum lustre_sec_part rq_sp_from; | |
1380 | ||
1381 | /* client/server security flags */ | |
1382 | unsigned int | |
1383 | rq_ctx_init:1, /* context initiation */ | |
1384 | rq_ctx_fini:1, /* context destroy */ | |
1385 | rq_bulk_read:1, /* request bulk read */ | |
1386 | rq_bulk_write:1, /* request bulk write */ | |
1387 | /* server authentication flags */ | |
1388 | rq_auth_gss:1, /* authenticated by gss */ | |
1389 | rq_auth_remote:1, /* authed as remote user */ | |
1390 | rq_auth_usr_root:1, /* authed as root */ | |
1391 | rq_auth_usr_mdt:1, /* authed as mdt */ | |
1392 | rq_auth_usr_ost:1, /* authed as ost */ | |
1393 | /* security tfm flags */ | |
1394 | rq_pack_udesc:1, | |
1395 | rq_pack_bulk:1, | |
1396 | /* doesn't expect reply FIXME */ | |
1397 | rq_no_reply:1, | |
1398 | rq_pill_init:1; /* pill initialized */ | |
1399 | ||
1400 | uid_t rq_auth_uid; /* authed uid */ | |
1401 | uid_t rq_auth_mapped_uid; /* authed uid mapped to */ | |
1402 | ||
1403 | /* (server side), pointed directly into req buffer */ | |
1404 | struct ptlrpc_user_desc *rq_user_desc; | |
1405 | ||
1406 | /* various buffer pointers */ | |
1407 | struct lustre_msg *rq_reqbuf; /* req wrapper */ | |
1408 | char *rq_repbuf; /* rep buffer */ | |
1409 | struct lustre_msg *rq_repdata; /* rep wrapper msg */ | |
1410 | struct lustre_msg *rq_clrbuf; /* only in priv mode */ | |
1411 | int rq_reqbuf_len; /* req wrapper buf len */ | |
1412 | int rq_reqdata_len; /* req wrapper msg len */ | |
1413 | int rq_repbuf_len; /* rep buffer len */ | |
1414 | int rq_repdata_len; /* rep wrapper msg len */ | |
1415 | int rq_clrbuf_len; /* only in priv mode */ | |
1416 | int rq_clrdata_len; /* only in priv mode */ | |
1417 | ||
1418 | /** early replies go to offset 0, regular replies go after that */ | |
1419 | unsigned int rq_reply_off; | |
1420 | ||
1421 | /** @} */ | |
1422 | ||
1423 | /** Fields that help to see if request and reply were swabbed or not */ | |
1424 | __u32 rq_req_swab_mask; | |
1425 | __u32 rq_rep_swab_mask; | |
1426 | ||
1427 | /** What was import generation when this request was sent */ | |
1428 | int rq_import_generation; | |
1429 | enum lustre_imp_state rq_send_state; | |
1430 | ||
1431 | /** how many early replies (for stats) */ | |
1432 | int rq_early_count; | |
1433 | ||
1434 | /** client+server request */ | |
1435 | lnet_handle_md_t rq_req_md_h; | |
1436 | struct ptlrpc_cb_id rq_req_cbid; | |
1437 | /** optional time limit for send attempts */ | |
b2d201bd | 1438 | long rq_delay_limit; |
d7e09d03 | 1439 | /** time request was first queued */ |
a649ad1d | 1440 | unsigned long rq_queued_time; |
d7e09d03 PT |
1441 | |
1442 | /* server-side... */ | |
1443 | /** request arrival time */ | |
219e6de6 | 1444 | struct timespec64 rq_arrival_time; |
d7e09d03 PT |
1445 | /** separated reply state */ |
1446 | struct ptlrpc_reply_state *rq_reply_state; | |
1447 | /** incoming request buffer */ | |
1448 | struct ptlrpc_request_buffer_desc *rq_rqbd; | |
1449 | ||
1450 | /** client-only incoming reply */ | |
1451 | lnet_handle_md_t rq_reply_md_h; | |
1452 | wait_queue_head_t rq_reply_waitq; | |
1453 | struct ptlrpc_cb_id rq_reply_cbid; | |
1454 | ||
1455 | /** our LNet NID */ | |
1456 | lnet_nid_t rq_self; | |
1457 | /** Peer description (the other side) */ | |
1458 | lnet_process_id_t rq_peer; | |
1459 | /** Server-side, export on which request was received */ | |
1460 | struct obd_export *rq_export; | |
1461 | /** Client side, import where request is being sent */ | |
1462 | struct obd_import *rq_import; | |
1463 | ||
1464 | /** Replay callback, called after request is replayed at recovery */ | |
1465 | void (*rq_replay_cb)(struct ptlrpc_request *); | |
1466 | /** | |
1467 | * Commit callback, called when request is committed and about to be | |
1468 | * freed. | |
1469 | */ | |
1470 | void (*rq_commit_cb)(struct ptlrpc_request *); | |
1471 | /** Opaq data for replay and commit callbacks. */ | |
1472 | void *rq_cb_data; | |
1473 | ||
1474 | /** For bulk requests on client only: bulk descriptor */ | |
1475 | struct ptlrpc_bulk_desc *rq_bulk; | |
1476 | ||
1477 | /** client outgoing req */ | |
1478 | /** | |
1479 | * when request/reply sent (secs), or time when request should be sent | |
1480 | */ | |
219e6de6 | 1481 | time64_t rq_sent; |
d7e09d03 | 1482 | /** time for request really sent out */ |
219e6de6 | 1483 | time64_t rq_real_sent; |
d7e09d03 PT |
1484 | |
1485 | /** when request must finish. volatile | |
1486 | * so that servers' early reply updates to the deadline aren't | |
c56e256d OD |
1487 | * kept in per-cpu cache |
1488 | */ | |
219e6de6 | 1489 | volatile time64_t rq_deadline; |
d7e09d03 | 1490 | /** when req reply unlink must finish. */ |
219e6de6 | 1491 | time64_t rq_reply_deadline; |
d7e09d03 | 1492 | /** when req bulk unlink must finish. */ |
219e6de6 | 1493 | time64_t rq_bulk_deadline; |
d7e09d03 PT |
1494 | /** |
1495 | * service time estimate (secs) | |
1496 | * If the requestsis not served by this time, it is marked as timed out. | |
1497 | */ | |
1498 | int rq_timeout; | |
1499 | ||
1500 | /** Multi-rpc bits */ | |
1501 | /** Per-request waitq introduced by bug 21938 for recovery waiting */ | |
1502 | wait_queue_head_t rq_set_waitq; | |
1503 | /** Link item for request set lists */ | |
1504 | struct list_head rq_set_chain; | |
1505 | /** Link back to the request set */ | |
1506 | struct ptlrpc_request_set *rq_set; | |
1507 | /** Async completion handler, called when reply is received */ | |
1508 | ptlrpc_interpterer_t rq_interpret_reply; | |
1509 | /** Async completion context */ | |
1510 | union ptlrpc_async_args rq_async_args; | |
1511 | ||
1512 | /** Pool if request is from preallocated list */ | |
1513 | struct ptlrpc_request_pool *rq_pool; | |
1514 | ||
1515 | struct lu_context rq_session; | |
1516 | struct lu_context rq_recov_session; | |
1517 | ||
1518 | /** request format description */ | |
1519 | struct req_capsule rq_pill; | |
1520 | }; | |
1521 | ||
1522 | /** | |
1523 | * Call completion handler for rpc if any, return it's status or original | |
1524 | * rc if there was no handler defined for this request. | |
1525 | */ | |
1526 | static inline int ptlrpc_req_interpret(const struct lu_env *env, | |
1527 | struct ptlrpc_request *req, int rc) | |
1528 | { | |
d2a13989 | 1529 | if (req->rq_interpret_reply) { |
d7e09d03 PT |
1530 | req->rq_status = req->rq_interpret_reply(env, req, |
1531 | &req->rq_async_args, | |
1532 | rc); | |
1533 | return req->rq_status; | |
1534 | } | |
1535 | return rc; | |
1536 | } | |
1537 | ||
d7e09d03 PT |
1538 | /* |
1539 | * Can the request be moved from the regular NRS head to the high-priority NRS | |
1540 | * head (of the same PTLRPC service partition), if any? | |
1541 | * | |
1542 | * For a reliable result, this should be checked under svcpt->scp_req lock. | |
1543 | */ | |
1544 | static inline bool ptlrpc_nrs_req_can_move(struct ptlrpc_request *req) | |
1545 | { | |
1546 | struct ptlrpc_nrs_request *nrq = &req->rq_nrq; | |
1547 | ||
1548 | /** | |
1549 | * LU-898: Check ptlrpc_nrs_request::nr_enqueued to make sure the | |
1550 | * request has been enqueued first, and ptlrpc_nrs_request::nr_started | |
1551 | * to make sure it has not been scheduled yet (analogous to previous | |
1552 | * (non-NRS) checking of !list_empty(&ptlrpc_request::rq_list). | |
1553 | */ | |
1554 | return nrq->nr_enqueued && !nrq->nr_started && !req->rq_hp; | |
1555 | } | |
c9f6bb96 | 1556 | |
d7e09d03 PT |
1557 | /** @} nrs */ |
1558 | ||
1559 | /** | |
1560 | * Returns 1 if request buffer at offset \a index was already swabbed | |
1561 | */ | |
1562 | static inline int lustre_req_swabbed(struct ptlrpc_request *req, int index) | |
1563 | { | |
1564 | LASSERT(index < sizeof(req->rq_req_swab_mask) * 8); | |
1565 | return req->rq_req_swab_mask & (1 << index); | |
1566 | } | |
1567 | ||
1568 | /** | |
1569 | * Returns 1 if request reply buffer at offset \a index was already swabbed | |
1570 | */ | |
1571 | static inline int lustre_rep_swabbed(struct ptlrpc_request *req, int index) | |
1572 | { | |
1573 | LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8); | |
1574 | return req->rq_rep_swab_mask & (1 << index); | |
1575 | } | |
1576 | ||
1577 | /** | |
1578 | * Returns 1 if request needs to be swabbed into local cpu byteorder | |
1579 | */ | |
1580 | static inline int ptlrpc_req_need_swab(struct ptlrpc_request *req) | |
1581 | { | |
1582 | return lustre_req_swabbed(req, MSG_PTLRPC_HEADER_OFF); | |
1583 | } | |
1584 | ||
1585 | /** | |
1586 | * Returns 1 if request reply needs to be swabbed into local cpu byteorder | |
1587 | */ | |
1588 | static inline int ptlrpc_rep_need_swab(struct ptlrpc_request *req) | |
1589 | { | |
1590 | return lustre_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF); | |
1591 | } | |
1592 | ||
1593 | /** | |
1594 | * Mark request buffer at offset \a index that it was already swabbed | |
1595 | */ | |
1596 | static inline void lustre_set_req_swabbed(struct ptlrpc_request *req, int index) | |
1597 | { | |
1598 | LASSERT(index < sizeof(req->rq_req_swab_mask) * 8); | |
1599 | LASSERT((req->rq_req_swab_mask & (1 << index)) == 0); | |
1600 | req->rq_req_swab_mask |= 1 << index; | |
1601 | } | |
1602 | ||
1603 | /** | |
1604 | * Mark request reply buffer at offset \a index that it was already swabbed | |
1605 | */ | |
1606 | static inline void lustre_set_rep_swabbed(struct ptlrpc_request *req, int index) | |
1607 | { | |
1608 | LASSERT(index < sizeof(req->rq_rep_swab_mask) * 8); | |
1609 | LASSERT((req->rq_rep_swab_mask & (1 << index)) == 0); | |
1610 | req->rq_rep_swab_mask |= 1 << index; | |
1611 | } | |
1612 | ||
1613 | /** | |
1614 | * Convert numerical request phase value \a phase into text string description | |
1615 | */ | |
1616 | static inline const char * | |
1617 | ptlrpc_phase2str(enum rq_phase phase) | |
1618 | { | |
1619 | switch (phase) { | |
1620 | case RQ_PHASE_NEW: | |
1621 | return "New"; | |
1622 | case RQ_PHASE_RPC: | |
1623 | return "Rpc"; | |
1624 | case RQ_PHASE_BULK: | |
1625 | return "Bulk"; | |
1626 | case RQ_PHASE_INTERPRET: | |
1627 | return "Interpret"; | |
1628 | case RQ_PHASE_COMPLETE: | |
1629 | return "Complete"; | |
1630 | case RQ_PHASE_UNREGISTERING: | |
1631 | return "Unregistering"; | |
1632 | default: | |
1633 | return "?Phase?"; | |
1634 | } | |
1635 | } | |
1636 | ||
1637 | /** | |
1638 | * Convert numerical request phase of the request \a req into text stringi | |
1639 | * description | |
1640 | */ | |
1641 | static inline const char * | |
1642 | ptlrpc_rqphase2str(struct ptlrpc_request *req) | |
1643 | { | |
1644 | return ptlrpc_phase2str(req->rq_phase); | |
1645 | } | |
1646 | ||
1647 | /** | |
1648 | * Debugging functions and helpers to print request structure into debug log | |
1649 | * @{ | |
1650 | */ | |
1651 | /* Spare the preprocessor, spoil the bugs. */ | |
1652 | #define FLAG(field, str) (field ? str : "") | |
1653 | ||
1654 | /** Convert bit flags into a string */ | |
1655 | #define DEBUG_REQ_FLAGS(req) \ | |
1656 | ptlrpc_rqphase2str(req), \ | |
1657 | FLAG(req->rq_intr, "I"), FLAG(req->rq_replied, "R"), \ | |
1658 | FLAG(req->rq_err, "E"), \ | |
1659 | FLAG(req->rq_timedout, "X") /* eXpired */, FLAG(req->rq_resend, "S"), \ | |
1660 | FLAG(req->rq_restart, "T"), FLAG(req->rq_replay, "P"), \ | |
1661 | FLAG(req->rq_no_resend, "N"), \ | |
1662 | FLAG(req->rq_waiting, "W"), \ | |
1663 | FLAG(req->rq_wait_ctx, "C"), FLAG(req->rq_hp, "H"), \ | |
1664 | FLAG(req->rq_committed, "M") | |
1665 | ||
1666 | #define REQ_FLAGS_FMT "%s:%s%s%s%s%s%s%s%s%s%s%s%s" | |
1667 | ||
1668 | void _debug_req(struct ptlrpc_request *req, | |
1669 | struct libcfs_debug_msg_data *data, const char *fmt, ...) | |
70837c12 | 1670 | __printf(3, 4); |
d7e09d03 PT |
1671 | |
1672 | /** | |
17891183 | 1673 | * Helper that decides if we need to print request according to current debug |
d7e09d03 PT |
1674 | * level settings |
1675 | */ | |
1676 | #define debug_req(msgdata, mask, cdls, req, fmt, a...) \ | |
1677 | do { \ | |
1678 | CFS_CHECK_STACK(msgdata, mask, cdls); \ | |
1679 | \ | |
1680 | if (((mask) & D_CANTMASK) != 0 || \ | |
1681 | ((libcfs_debug & (mask)) != 0 && \ | |
1682 | (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \ | |
1683 | _debug_req((req), msgdata, fmt, ##a); \ | |
a58a38ac | 1684 | } while (0) |
d7e09d03 PT |
1685 | |
1686 | /** | |
17891183 | 1687 | * This is the debug print function you need to use to print request structure |
d7e09d03 | 1688 | * content into lustre debug log. |
c56e256d OD |
1689 | * for most callers (level is a constant) this is resolved at compile time |
1690 | */ | |
d7e09d03 PT |
1691 | #define DEBUG_REQ(level, req, fmt, args...) \ |
1692 | do { \ | |
1693 | if ((level) & (D_ERROR | D_WARNING)) { \ | |
a3ea59e0 | 1694 | static struct cfs_debug_limit_state cdls; \ |
d7e09d03 PT |
1695 | LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \ |
1696 | debug_req(&msgdata, level, &cdls, req, "@@@ "fmt" ", ## args);\ | |
1697 | } else { \ | |
1698 | LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \ | |
1699 | debug_req(&msgdata, level, NULL, req, "@@@ "fmt" ", ## args); \ | |
1700 | } \ | |
1701 | } while (0) | |
1702 | /** @} */ | |
1703 | ||
1704 | /** | |
1705 | * Structure that defines a single page of a bulk transfer | |
1706 | */ | |
1707 | struct ptlrpc_bulk_page { | |
1708 | /** Linkage to list of pages in a bulk */ | |
1709 | struct list_head bp_link; | |
1710 | /** | |
1711 | * Number of bytes in a page to transfer starting from \a bp_pageoffset | |
1712 | */ | |
1713 | int bp_buflen; | |
1714 | /** offset within a page */ | |
1715 | int bp_pageoffset; | |
1716 | /** The page itself */ | |
1717 | struct page *bp_page; | |
1718 | }; | |
1719 | ||
1720 | #define BULK_GET_SOURCE 0 | |
1721 | #define BULK_PUT_SINK 1 | |
1722 | #define BULK_GET_SINK 2 | |
1723 | #define BULK_PUT_SOURCE 3 | |
1724 | ||
1725 | /** | |
1726 | * Definition of bulk descriptor. | |
1727 | * Bulks are special "Two phase" RPCs where initial request message | |
1728 | * is sent first and it is followed bt a transfer (o receiving) of a large | |
1729 | * amount of data to be settled into pages referenced from the bulk descriptors. | |
1730 | * Bulks transfers (the actual data following the small requests) are done | |
1731 | * on separate LNet portals. | |
1732 | * In lustre we use bulk transfers for READ and WRITE transfers from/to OSTs. | |
1733 | * Another user is readpage for MDT. | |
1734 | */ | |
1735 | struct ptlrpc_bulk_desc { | |
1736 | /** completed with failure */ | |
1737 | unsigned long bd_failure:1; | |
1738 | /** {put,get}{source,sink} */ | |
1739 | unsigned long bd_type:2; | |
1740 | /** client side */ | |
1741 | unsigned long bd_registered:1; | |
1742 | /** For serialization with callback */ | |
1743 | spinlock_t bd_lock; | |
1744 | /** Import generation when request for this bulk was sent */ | |
1745 | int bd_import_generation; | |
1746 | /** LNet portal for this bulk */ | |
1747 | __u32 bd_portal; | |
1748 | /** Server side - export this bulk created for */ | |
1749 | struct obd_export *bd_export; | |
1750 | /** Client side - import this bulk was sent on */ | |
1751 | struct obd_import *bd_import; | |
1752 | /** Back pointer to the request */ | |
1753 | struct ptlrpc_request *bd_req; | |
1754 | wait_queue_head_t bd_waitq; /* server side only WQ */ | |
1755 | int bd_iov_count; /* # entries in bd_iov */ | |
1756 | int bd_max_iov; /* allocated size of bd_iov */ | |
1757 | int bd_nob; /* # bytes covered */ | |
1758 | int bd_nob_transferred; /* # bytes GOT/PUT */ | |
1759 | ||
1760 | __u64 bd_last_xid; | |
1761 | ||
1762 | struct ptlrpc_cb_id bd_cbid; /* network callback info */ | |
1763 | lnet_nid_t bd_sender; /* stash event::sender */ | |
1764 | int bd_md_count; /* # valid entries in bd_mds */ | |
1765 | int bd_md_max_brw; /* max entries in bd_mds */ | |
1766 | /** array of associated MDs */ | |
1767 | lnet_handle_md_t bd_mds[PTLRPC_BULK_OPS_COUNT]; | |
1768 | ||
1769 | /* | |
1770 | * encrypt iov, size is either 0 or bd_iov_count. | |
1771 | */ | |
1772 | lnet_kiov_t *bd_enc_iov; | |
1773 | ||
1774 | lnet_kiov_t bd_iov[0]; | |
1775 | }; | |
1776 | ||
1777 | enum { | |
1778 | SVC_STOPPED = 1 << 0, | |
1779 | SVC_STOPPING = 1 << 1, | |
1780 | SVC_STARTING = 1 << 2, | |
1781 | SVC_RUNNING = 1 << 3, | |
1782 | SVC_EVENT = 1 << 4, | |
1783 | SVC_SIGNAL = 1 << 5, | |
1784 | }; | |
1785 | ||
1786 | #define PTLRPC_THR_NAME_LEN 32 | |
1787 | /** | |
1788 | * Definition of server service thread structure | |
1789 | */ | |
1790 | struct ptlrpc_thread { | |
1791 | /** | |
1792 | * List of active threads in svc->srv_threads | |
1793 | */ | |
1794 | struct list_head t_link; | |
1795 | /** | |
1796 | * thread-private data (preallocated memory) | |
1797 | */ | |
1798 | void *t_data; | |
1799 | __u32 t_flags; | |
1800 | /** | |
1801 | * service thread index, from ptlrpc_start_threads | |
1802 | */ | |
1803 | unsigned int t_id; | |
1804 | /** | |
1805 | * service thread pid | |
1806 | */ | |
1807 | pid_t t_pid; | |
1808 | /** | |
1809 | * put watchdog in the structure per thread b=14840 | |
5d4450c4 PT |
1810 | * |
1811 | * Lustre watchdog is removed for client in the hope | |
1812 | * of a generic watchdog can be merged in kernel. | |
1813 | * When that happens, we should add below back. | |
1814 | * | |
1815 | * struct lc_watchdog *t_watchdog; | |
d7e09d03 | 1816 | */ |
d7e09d03 PT |
1817 | /** |
1818 | * the svc this thread belonged to b=18582 | |
1819 | */ | |
1820 | struct ptlrpc_service_part *t_svcpt; | |
1821 | wait_queue_head_t t_ctl_waitq; | |
1822 | struct lu_env *t_env; | |
1823 | char t_name[PTLRPC_THR_NAME_LEN]; | |
1824 | }; | |
1825 | ||
1826 | static inline int thread_is_init(struct ptlrpc_thread *thread) | |
1827 | { | |
1828 | return thread->t_flags == 0; | |
1829 | } | |
1830 | ||
1831 | static inline int thread_is_stopped(struct ptlrpc_thread *thread) | |
1832 | { | |
1833 | return !!(thread->t_flags & SVC_STOPPED); | |
1834 | } | |
1835 | ||
1836 | static inline int thread_is_stopping(struct ptlrpc_thread *thread) | |
1837 | { | |
1838 | return !!(thread->t_flags & SVC_STOPPING); | |
1839 | } | |
1840 | ||
1841 | static inline int thread_is_starting(struct ptlrpc_thread *thread) | |
1842 | { | |
1843 | return !!(thread->t_flags & SVC_STARTING); | |
1844 | } | |
1845 | ||
1846 | static inline int thread_is_running(struct ptlrpc_thread *thread) | |
1847 | { | |
1848 | return !!(thread->t_flags & SVC_RUNNING); | |
1849 | } | |
1850 | ||
1851 | static inline int thread_is_event(struct ptlrpc_thread *thread) | |
1852 | { | |
1853 | return !!(thread->t_flags & SVC_EVENT); | |
1854 | } | |
1855 | ||
1856 | static inline int thread_is_signal(struct ptlrpc_thread *thread) | |
1857 | { | |
1858 | return !!(thread->t_flags & SVC_SIGNAL); | |
1859 | } | |
1860 | ||
1861 | static inline void thread_clear_flags(struct ptlrpc_thread *thread, __u32 flags) | |
1862 | { | |
1863 | thread->t_flags &= ~flags; | |
1864 | } | |
1865 | ||
1866 | static inline void thread_set_flags(struct ptlrpc_thread *thread, __u32 flags) | |
1867 | { | |
1868 | thread->t_flags = flags; | |
1869 | } | |
1870 | ||
1871 | static inline void thread_add_flags(struct ptlrpc_thread *thread, __u32 flags) | |
1872 | { | |
1873 | thread->t_flags |= flags; | |
1874 | } | |
1875 | ||
1876 | static inline int thread_test_and_clear_flags(struct ptlrpc_thread *thread, | |
1877 | __u32 flags) | |
1878 | { | |
1879 | if (thread->t_flags & flags) { | |
1880 | thread->t_flags &= ~flags; | |
1881 | return 1; | |
1882 | } | |
1883 | return 0; | |
1884 | } | |
1885 | ||
1886 | /** | |
1887 | * Request buffer descriptor structure. | |
1888 | * This is a structure that contains one posted request buffer for service. | |
1889 | * Once data land into a buffer, event callback creates actual request and | |
1890 | * notifies wakes one of the service threads to process new incoming request. | |
1891 | * More than one request can fit into the buffer. | |
1892 | */ | |
1893 | struct ptlrpc_request_buffer_desc { | |
1894 | /** Link item for rqbds on a service */ | |
1895 | struct list_head rqbd_list; | |
1896 | /** History of requests for this buffer */ | |
1897 | struct list_head rqbd_reqs; | |
1898 | /** Back pointer to service for which this buffer is registered */ | |
1899 | struct ptlrpc_service_part *rqbd_svcpt; | |
1900 | /** LNet descriptor */ | |
1901 | lnet_handle_md_t rqbd_md_h; | |
1902 | int rqbd_refcount; | |
1903 | /** The buffer itself */ | |
1904 | char *rqbd_buffer; | |
1905 | struct ptlrpc_cb_id rqbd_cbid; | |
1906 | /** | |
1907 | * This "embedded" request structure is only used for the | |
1908 | * last request to fit into the buffer | |
1909 | */ | |
1910 | struct ptlrpc_request rqbd_req; | |
1911 | }; | |
1912 | ||
1913 | typedef int (*svc_handler_t)(struct ptlrpc_request *req); | |
1914 | ||
1915 | struct ptlrpc_service_ops { | |
1916 | /** | |
1917 | * if non-NULL called during thread creation (ptlrpc_start_thread()) | |
1918 | * to initialize service specific per-thread state. | |
1919 | */ | |
1920 | int (*so_thr_init)(struct ptlrpc_thread *thr); | |
1921 | /** | |
1922 | * if non-NULL called during thread shutdown (ptlrpc_main()) to | |
1923 | * destruct state created by ->srv_init(). | |
1924 | */ | |
1925 | void (*so_thr_done)(struct ptlrpc_thread *thr); | |
1926 | /** | |
1927 | * Handler function for incoming requests for this service | |
1928 | */ | |
1929 | int (*so_req_handler)(struct ptlrpc_request *req); | |
1930 | /** | |
1931 | * function to determine priority of the request, it's called | |
1932 | * on every new request | |
1933 | */ | |
1934 | int (*so_hpreq_handler)(struct ptlrpc_request *); | |
1935 | /** | |
1936 | * service-specific print fn | |
1937 | */ | |
1938 | void (*so_req_printer)(void *, struct ptlrpc_request *); | |
1939 | }; | |
1940 | ||
1941 | #ifndef __cfs_cacheline_aligned | |
1942 | /* NB: put it here for reducing patche dependence */ | |
1943 | # define __cfs_cacheline_aligned | |
1944 | #endif | |
1945 | ||
1946 | /** | |
1947 | * How many high priority requests to serve before serving one normal | |
1948 | * priority request | |
1949 | */ | |
1950 | #define PTLRPC_SVC_HP_RATIO 10 | |
1951 | ||
1952 | /** | |
1953 | * Definition of PortalRPC service. | |
1954 | * The service is listening on a particular portal (like tcp port) | |
1955 | * and perform actions for a specific server like IO service for OST | |
1956 | * or general metadata service for MDS. | |
1957 | */ | |
1958 | struct ptlrpc_service { | |
f6e42a40 | 1959 | /** serialize sysfs operations */ |
d7e09d03 PT |
1960 | spinlock_t srv_lock; |
1961 | /** most often accessed fields */ | |
1962 | /** chain thru all services */ | |
1963 | struct list_head srv_list; | |
1964 | /** service operations table */ | |
1965 | struct ptlrpc_service_ops srv_ops; | |
1966 | /** only statically allocated strings here; we don't clean them */ | |
1967 | char *srv_name; | |
1968 | /** only statically allocated strings here; we don't clean them */ | |
1969 | char *srv_thread_name; | |
1970 | /** service thread list */ | |
1971 | struct list_head srv_threads; | |
1972 | /** threads # should be created for each partition on initializing */ | |
1973 | int srv_nthrs_cpt_init; | |
1974 | /** limit of threads number for each partition */ | |
1975 | int srv_nthrs_cpt_limit; | |
700815d4 DE |
1976 | /** Root of debugfs dir tree for this service */ |
1977 | struct dentry *srv_debugfs_entry; | |
d7e09d03 PT |
1978 | /** Pointer to statistic data for this service */ |
1979 | struct lprocfs_stats *srv_stats; | |
1980 | /** # hp per lp reqs to handle */ | |
1981 | int srv_hpreq_ratio; | |
1982 | /** biggest request to receive */ | |
1983 | int srv_max_req_size; | |
1984 | /** biggest reply to send */ | |
1985 | int srv_max_reply_size; | |
1986 | /** size of individual buffers */ | |
1987 | int srv_buf_size; | |
1988 | /** # buffers to allocate in 1 group */ | |
1989 | int srv_nbuf_per_group; | |
1990 | /** Local portal on which to receive requests */ | |
1991 | __u32 srv_req_portal; | |
1992 | /** Portal on the client to send replies to */ | |
1993 | __u32 srv_rep_portal; | |
1994 | /** | |
1995 | * Tags for lu_context associated with this thread, see struct | |
1996 | * lu_context. | |
1997 | */ | |
1998 | __u32 srv_ctx_tags; | |
1999 | /** soft watchdog timeout multiplier */ | |
2000 | int srv_watchdog_factor; | |
2001 | /** under unregister_service */ | |
2002 | unsigned srv_is_stopping:1; | |
2003 | ||
2004 | /** max # request buffers in history per partition */ | |
2005 | int srv_hist_nrqbds_cpt_max; | |
2006 | /** number of CPTs this service bound on */ | |
2007 | int srv_ncpts; | |
2008 | /** CPTs array this service bound on */ | |
2009 | __u32 *srv_cpts; | |
2010 | /** 2^srv_cptab_bits >= cfs_cpt_numbert(srv_cptable) */ | |
2011 | int srv_cpt_bits; | |
2012 | /** CPT table this service is running over */ | |
2013 | struct cfs_cpt_table *srv_cptable; | |
328676f8 OD |
2014 | |
2015 | /* sysfs object */ | |
2016 | struct kobject srv_kobj; | |
2017 | struct completion srv_kobj_unregister; | |
d7e09d03 PT |
2018 | /** |
2019 | * partition data for ptlrpc service | |
2020 | */ | |
2021 | struct ptlrpc_service_part *srv_parts[0]; | |
2022 | }; | |
2023 | ||
2024 | /** | |
2025 | * Definition of PortalRPC service partition data. | |
2026 | * Although a service only has one instance of it right now, but we | |
2027 | * will have multiple instances very soon (instance per CPT). | |
2028 | * | |
2029 | * it has four locks: | |
2030 | * \a scp_lock | |
2031 | * serialize operations on rqbd and requests waiting for preprocess | |
2032 | * \a scp_req_lock | |
2033 | * serialize operations active requests sent to this portal | |
2034 | * \a scp_at_lock | |
2035 | * serialize adaptive timeout stuff | |
2036 | * \a scp_rep_lock | |
2037 | * serialize operations on RS list (reply states) | |
2038 | * | |
2039 | * We don't have any use-case to take two or more locks at the same time | |
2040 | * for now, so there is no lock order issue. | |
2041 | */ | |
2042 | struct ptlrpc_service_part { | |
2043 | /** back reference to owner */ | |
2044 | struct ptlrpc_service *scp_service __cfs_cacheline_aligned; | |
2045 | /* CPT id, reserved */ | |
2046 | int scp_cpt; | |
2047 | /** always increasing number */ | |
2048 | int scp_thr_nextid; | |
2049 | /** # of starting threads */ | |
2050 | int scp_nthrs_starting; | |
2051 | /** # of stopping threads, reserved for shrinking threads */ | |
2052 | int scp_nthrs_stopping; | |
2053 | /** # running threads */ | |
2054 | int scp_nthrs_running; | |
2055 | /** service threads list */ | |
2056 | struct list_head scp_threads; | |
2057 | ||
2058 | /** | |
2059 | * serialize the following fields, used for protecting | |
2060 | * rqbd list and incoming requests waiting for preprocess, | |
2061 | * threads starting & stopping are also protected by this lock. | |
2062 | */ | |
b2952d62 | 2063 | spinlock_t scp_lock __cfs_cacheline_aligned; |
d7e09d03 PT |
2064 | /** total # req buffer descs allocated */ |
2065 | int scp_nrqbds_total; | |
2066 | /** # posted request buffers for receiving */ | |
2067 | int scp_nrqbds_posted; | |
2068 | /** in progress of allocating rqbd */ | |
2069 | int scp_rqbd_allocating; | |
2070 | /** # incoming reqs */ | |
2071 | int scp_nreqs_incoming; | |
2072 | /** request buffers to be reposted */ | |
2073 | struct list_head scp_rqbd_idle; | |
2074 | /** req buffers receiving */ | |
2075 | struct list_head scp_rqbd_posted; | |
2076 | /** incoming reqs */ | |
2077 | struct list_head scp_req_incoming; | |
2078 | /** timeout before re-posting reqs, in tick */ | |
b2d201bd | 2079 | long scp_rqbd_timeout; |
d7e09d03 PT |
2080 | /** |
2081 | * all threads sleep on this. This wait-queue is signalled when new | |
2082 | * incoming request arrives and when difficult reply has to be handled. | |
2083 | */ | |
2084 | wait_queue_head_t scp_waitq; | |
2085 | ||
2086 | /** request history */ | |
2087 | struct list_head scp_hist_reqs; | |
2088 | /** request buffer history */ | |
2089 | struct list_head scp_hist_rqbds; | |
2090 | /** # request buffers in history */ | |
2091 | int scp_hist_nrqbds; | |
2092 | /** sequence number for request */ | |
2093 | __u64 scp_hist_seq; | |
2094 | /** highest seq culled from history */ | |
2095 | __u64 scp_hist_seq_culled; | |
2096 | ||
2097 | /** | |
2098 | * serialize the following fields, used for processing requests | |
2099 | * sent to this portal | |
2100 | */ | |
2101 | spinlock_t scp_req_lock __cfs_cacheline_aligned; | |
2102 | /** # reqs in either of the NRS heads below */ | |
2103 | /** # reqs being served */ | |
2104 | int scp_nreqs_active; | |
2105 | /** # HPreqs being served */ | |
2106 | int scp_nhreqs_active; | |
2107 | /** # hp requests handled */ | |
2108 | int scp_hreq_count; | |
2109 | ||
2110 | /** NRS head for regular requests */ | |
2111 | struct ptlrpc_nrs scp_nrs_reg; | |
2112 | /** NRS head for HP requests; this is only valid for services that can | |
c56e256d OD |
2113 | * handle HP requests |
2114 | */ | |
d7e09d03 PT |
2115 | struct ptlrpc_nrs *scp_nrs_hp; |
2116 | ||
2117 | /** AT stuff */ | |
2118 | /** @{ */ | |
2119 | /** | |
2120 | * serialize the following fields, used for changes on | |
2121 | * adaptive timeout | |
2122 | */ | |
2123 | spinlock_t scp_at_lock __cfs_cacheline_aligned; | |
2124 | /** estimated rpc service time */ | |
2125 | struct adaptive_timeout scp_at_estimate; | |
2126 | /** reqs waiting for replies */ | |
2127 | struct ptlrpc_at_array scp_at_array; | |
2128 | /** early reply timer */ | |
54319351 | 2129 | struct timer_list scp_at_timer; |
d7e09d03 | 2130 | /** debug */ |
a649ad1d | 2131 | unsigned long scp_at_checktime; |
d7e09d03 PT |
2132 | /** check early replies */ |
2133 | unsigned scp_at_check; | |
2134 | /** @} */ | |
2135 | ||
2136 | /** | |
2137 | * serialize the following fields, used for processing | |
2138 | * replies for this portal | |
2139 | */ | |
2140 | spinlock_t scp_rep_lock __cfs_cacheline_aligned; | |
2141 | /** all the active replies */ | |
2142 | struct list_head scp_rep_active; | |
2143 | /** List of free reply_states */ | |
2144 | struct list_head scp_rep_idle; | |
2145 | /** waitq to run, when adding stuff to srv_free_rs_list */ | |
2146 | wait_queue_head_t scp_rep_waitq; | |
2147 | /** # 'difficult' replies */ | |
2148 | atomic_t scp_nreps_difficult; | |
2149 | }; | |
2150 | ||
2151 | #define ptlrpc_service_for_each_part(part, i, svc) \ | |
2152 | for (i = 0; \ | |
2153 | i < (svc)->srv_ncpts && \ | |
d2a13989 OD |
2154 | (svc)->srv_parts && \ |
2155 | ((part) = (svc)->srv_parts[i]); i++) | |
d7e09d03 PT |
2156 | |
2157 | /** | |
2158 | * Declaration of ptlrpcd control structure | |
2159 | */ | |
2160 | struct ptlrpcd_ctl { | |
2161 | /** | |
2162 | * Ptlrpc thread control flags (LIOD_START, LIOD_STOP, LIOD_FORCE) | |
2163 | */ | |
2164 | unsigned long pc_flags; | |
2165 | /** | |
2166 | * Thread lock protecting structure fields. | |
2167 | */ | |
2168 | spinlock_t pc_lock; | |
2169 | /** | |
2170 | * Start completion. | |
2171 | */ | |
2172 | struct completion pc_starting; | |
2173 | /** | |
2174 | * Stop completion. | |
2175 | */ | |
2176 | struct completion pc_finishing; | |
2177 | /** | |
2178 | * Thread requests set. | |
2179 | */ | |
2180 | struct ptlrpc_request_set *pc_set; | |
2181 | /** | |
f72f130d | 2182 | * Thread name used in kthread_run() |
d7e09d03 PT |
2183 | */ |
2184 | char pc_name[16]; | |
2185 | /** | |
2186 | * Environment for request interpreters to run in. | |
2187 | */ | |
2188 | struct lu_env pc_env; | |
2189 | /** | |
c5c4c6fa | 2190 | * CPT the thread is bound on. |
d7e09d03 | 2191 | */ |
c5c4c6fa | 2192 | int pc_cpt; |
d7e09d03 | 2193 | /** |
c5c4c6fa | 2194 | * Index of ptlrpcd thread in the array. |
d7e09d03 | 2195 | */ |
c5c4c6fa | 2196 | int pc_index; |
d7e09d03 PT |
2197 | /** |
2198 | * Pointer to the array of partners' ptlrpcd_ctl structure. | |
2199 | */ | |
2200 | struct ptlrpcd_ctl **pc_partners; | |
c5c4c6fa OW |
2201 | /** |
2202 | * Number of the ptlrpcd's partners. | |
2203 | */ | |
2204 | int pc_npartners; | |
d7e09d03 PT |
2205 | /** |
2206 | * Record the partner index to be processed next. | |
2207 | */ | |
2208 | int pc_cursor; | |
c5c4c6fa OW |
2209 | /** |
2210 | * Error code if the thread failed to fully start. | |
2211 | */ | |
2212 | int pc_error; | |
d7e09d03 PT |
2213 | }; |
2214 | ||
2215 | /* Bits for pc_flags */ | |
2216 | enum ptlrpcd_ctl_flags { | |
2217 | /** | |
2218 | * Ptlrpc thread start flag. | |
2219 | */ | |
2220 | LIOD_START = 1 << 0, | |
2221 | /** | |
2222 | * Ptlrpc thread stop flag. | |
2223 | */ | |
2224 | LIOD_STOP = 1 << 1, | |
2225 | /** | |
2226 | * Ptlrpc thread force flag (only stop force so far). | |
2227 | * This will cause aborting any inflight rpcs handled | |
2228 | * by thread if LIOD_STOP is specified. | |
2229 | */ | |
2230 | LIOD_FORCE = 1 << 2, | |
2231 | /** | |
2232 | * This is a recovery ptlrpc thread. | |
2233 | */ | |
2234 | LIOD_RECOVERY = 1 << 3, | |
d7e09d03 PT |
2235 | }; |
2236 | ||
2237 | /** | |
2238 | * \addtogroup nrs | |
2239 | * @{ | |
2240 | * | |
2241 | * Service compatibility function; the policy is compatible with all services. | |
2242 | * | |
2243 | * \param[in] svc The service the policy is attempting to register with. | |
2244 | * \param[in] desc The policy descriptor | |
2245 | * | |
2246 | * \retval true The policy is compatible with the service | |
2247 | * | |
2248 | * \see ptlrpc_nrs_pol_desc::pd_compat() | |
2249 | */ | |
2250 | static inline bool nrs_policy_compat_all(const struct ptlrpc_service *svc, | |
2251 | const struct ptlrpc_nrs_pol_desc *desc) | |
2252 | { | |
2253 | return true; | |
2254 | } | |
2255 | ||
2256 | /** | |
2257 | * Service compatibility function; the policy is compatible with only a specific | |
2258 | * service which is identified by its human-readable name at | |
2259 | * ptlrpc_service::srv_name. | |
2260 | * | |
2261 | * \param[in] svc The service the policy is attempting to register with. | |
2262 | * \param[in] desc The policy descriptor | |
2263 | * | |
2264 | * \retval false The policy is not compatible with the service | |
2265 | * \retval true The policy is compatible with the service | |
2266 | * | |
2267 | * \see ptlrpc_nrs_pol_desc::pd_compat() | |
2268 | */ | |
2269 | static inline bool nrs_policy_compat_one(const struct ptlrpc_service *svc, | |
2270 | const struct ptlrpc_nrs_pol_desc *desc) | |
2271 | { | |
d7e09d03 PT |
2272 | return strcmp(svc->srv_name, desc->pd_compat_svc_name) == 0; |
2273 | } | |
2274 | ||
2275 | /** @} nrs */ | |
2276 | ||
2277 | /* ptlrpc/events.c */ | |
2278 | extern lnet_handle_eq_t ptlrpc_eq_h; | |
8150a97f JP |
2279 | int ptlrpc_uuid_to_peer(struct obd_uuid *uuid, |
2280 | lnet_process_id_t *peer, lnet_nid_t *self); | |
d7e09d03 PT |
2281 | /** |
2282 | * These callbacks are invoked by LNet when something happened to | |
2283 | * underlying buffer | |
2284 | * @{ | |
2285 | */ | |
8150a97f JP |
2286 | void request_out_callback(lnet_event_t *ev); |
2287 | void reply_in_callback(lnet_event_t *ev); | |
2288 | void client_bulk_callback(lnet_event_t *ev); | |
2289 | void request_in_callback(lnet_event_t *ev); | |
2290 | void reply_out_callback(lnet_event_t *ev); | |
d7e09d03 PT |
2291 | /** @} */ |
2292 | ||
2293 | /* ptlrpc/connection.c */ | |
2294 | struct ptlrpc_connection *ptlrpc_connection_get(lnet_process_id_t peer, | |
2295 | lnet_nid_t self, | |
2296 | struct obd_uuid *uuid); | |
2297 | int ptlrpc_connection_put(struct ptlrpc_connection *c); | |
2298 | struct ptlrpc_connection *ptlrpc_connection_addref(struct ptlrpc_connection *); | |
2299 | int ptlrpc_connection_init(void); | |
2300 | void ptlrpc_connection_fini(void); | |
d7e09d03 PT |
2301 | |
2302 | /* ptlrpc/niobuf.c */ | |
2303 | /** | |
2304 | * Actual interfacing with LNet to put/get/register/unregister stuff | |
2305 | * @{ | |
2306 | */ | |
2307 | ||
d7e09d03 PT |
2308 | int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async); |
2309 | ||
2310 | static inline int ptlrpc_client_bulk_active(struct ptlrpc_request *req) | |
2311 | { | |
2312 | struct ptlrpc_bulk_desc *desc; | |
2313 | int rc; | |
2314 | ||
d7e09d03 PT |
2315 | desc = req->rq_bulk; |
2316 | ||
2317 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) && | |
219e6de6 | 2318 | req->rq_bulk_deadline > ktime_get_real_seconds()) |
d7e09d03 PT |
2319 | return 1; |
2320 | ||
2321 | if (!desc) | |
2322 | return 0; | |
2323 | ||
2324 | spin_lock(&desc->bd_lock); | |
2325 | rc = desc->bd_md_count; | |
2326 | spin_unlock(&desc->bd_lock); | |
2327 | return rc; | |
2328 | } | |
2329 | ||
2330 | #define PTLRPC_REPLY_MAYBE_DIFFICULT 0x01 | |
2331 | #define PTLRPC_REPLY_EARLY 0x02 | |
2332 | int ptlrpc_send_reply(struct ptlrpc_request *req, int flags); | |
2333 | int ptlrpc_reply(struct ptlrpc_request *req); | |
2334 | int ptlrpc_send_error(struct ptlrpc_request *req, int difficult); | |
2335 | int ptlrpc_error(struct ptlrpc_request *req); | |
2336 | void ptlrpc_resend_req(struct ptlrpc_request *request); | |
2337 | int ptlrpc_at_get_net_latency(struct ptlrpc_request *req); | |
2338 | int ptl_send_rpc(struct ptlrpc_request *request, int noreply); | |
2339 | int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd); | |
2340 | /** @} */ | |
2341 | ||
2342 | /* ptlrpc/client.c */ | |
2343 | /** | |
2344 | * Client-side portals API. Everything to send requests, receive replies, | |
2345 | * request queues, request management, etc. | |
2346 | * @{ | |
2347 | */ | |
63d42578 HZ |
2348 | void ptlrpc_request_committed(struct ptlrpc_request *req, int force); |
2349 | ||
d7e09d03 PT |
2350 | void ptlrpc_init_client(int req_portal, int rep_portal, char *name, |
2351 | struct ptlrpc_client *); | |
d7e09d03 PT |
2352 | struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid); |
2353 | ||
2354 | int ptlrpc_queue_wait(struct ptlrpc_request *req); | |
2355 | int ptlrpc_replay_req(struct ptlrpc_request *req); | |
2356 | int ptlrpc_unregister_reply(struct ptlrpc_request *req, int async); | |
d7e09d03 | 2357 | void ptlrpc_abort_inflight(struct obd_import *imp); |
d7e09d03 PT |
2358 | void ptlrpc_abort_set(struct ptlrpc_request_set *set); |
2359 | ||
2360 | struct ptlrpc_request_set *ptlrpc_prep_set(void); | |
2361 | struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func, | |
2362 | void *arg); | |
d7e09d03 PT |
2363 | int ptlrpc_set_next_timeout(struct ptlrpc_request_set *); |
2364 | int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set); | |
2365 | int ptlrpc_set_wait(struct ptlrpc_request_set *); | |
2366 | int ptlrpc_expired_set(void *data); | |
2367 | void ptlrpc_interrupted_set(void *data); | |
2368 | void ptlrpc_mark_interrupted(struct ptlrpc_request *req); | |
2369 | void ptlrpc_set_destroy(struct ptlrpc_request_set *); | |
2370 | void ptlrpc_set_add_req(struct ptlrpc_request_set *, struct ptlrpc_request *); | |
2371 | void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, | |
2372 | struct ptlrpc_request *req); | |
2373 | ||
2374 | void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool); | |
aefd9d71 | 2375 | int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq); |
d7e09d03 PT |
2376 | |
2377 | struct ptlrpc_request_pool * | |
2378 | ptlrpc_init_rq_pool(int, int, | |
aefd9d71 | 2379 | int (*populate_pool)(struct ptlrpc_request_pool *, int)); |
d7e09d03 PT |
2380 | |
2381 | void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req); | |
2382 | struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp, | |
2383 | const struct req_format *format); | |
2384 | struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp, | |
10457d4b OD |
2385 | struct ptlrpc_request_pool *, |
2386 | const struct req_format *); | |
d7e09d03 PT |
2387 | void ptlrpc_request_free(struct ptlrpc_request *request); |
2388 | int ptlrpc_request_pack(struct ptlrpc_request *request, | |
2389 | __u32 version, int opcode); | |
10457d4b OD |
2390 | struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *, |
2391 | const struct req_format *, | |
2392 | __u32, int); | |
d7e09d03 PT |
2393 | int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, |
2394 | __u32 version, int opcode, char **bufs, | |
2395 | struct ptlrpc_cli_ctx *ctx); | |
d7e09d03 | 2396 | void ptlrpc_req_finished(struct ptlrpc_request *request); |
d7e09d03 PT |
2397 | struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req); |
2398 | struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, | |
2399 | unsigned npages, unsigned max_brw, | |
2400 | unsigned type, unsigned portal); | |
2401 | void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk, int pin); | |
2402 | static inline void ptlrpc_free_bulk_pin(struct ptlrpc_bulk_desc *bulk) | |
2403 | { | |
2404 | __ptlrpc_free_bulk(bulk, 1); | |
2405 | } | |
c9f6bb96 | 2406 | |
d7e09d03 PT |
2407 | static inline void ptlrpc_free_bulk_nopin(struct ptlrpc_bulk_desc *bulk) |
2408 | { | |
2409 | __ptlrpc_free_bulk(bulk, 0); | |
2410 | } | |
c9f6bb96 | 2411 | |
d7e09d03 PT |
2412 | void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc, |
2413 | struct page *page, int pageoffset, int len, int); | |
2414 | static inline void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc, | |
2415 | struct page *page, int pageoffset, | |
2416 | int len) | |
2417 | { | |
2418 | __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1); | |
2419 | } | |
2420 | ||
2421 | static inline void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc, | |
2422 | struct page *page, int pageoffset, | |
2423 | int len) | |
2424 | { | |
2425 | __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0); | |
2426 | } | |
2427 | ||
2428 | void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, | |
2429 | struct obd_import *imp); | |
2430 | __u64 ptlrpc_next_xid(void); | |
2431 | __u64 ptlrpc_sample_next_xid(void); | |
2432 | __u64 ptlrpc_req_xid(struct ptlrpc_request *request); | |
2433 | ||
2434 | /* Set of routines to run a function in ptlrpcd context */ | |
2435 | void *ptlrpcd_alloc_work(struct obd_import *imp, | |
2436 | int (*cb)(const struct lu_env *, void *), void *data); | |
2437 | void ptlrpcd_destroy_work(void *handler); | |
2438 | int ptlrpcd_queue_work(void *handler); | |
2439 | ||
2440 | /** @} */ | |
2441 | struct ptlrpc_service_buf_conf { | |
2442 | /* nbufs is buffers # to allocate when growing the pool */ | |
2443 | unsigned int bc_nbufs; | |
2444 | /* buffer size to post */ | |
2445 | unsigned int bc_buf_size; | |
2446 | /* portal to listed for requests on */ | |
2447 | unsigned int bc_req_portal; | |
2448 | /* portal of where to send replies to */ | |
2449 | unsigned int bc_rep_portal; | |
2450 | /* maximum request size to be accepted for this service */ | |
2451 | unsigned int bc_req_max_size; | |
2452 | /* maximum reply size this service can ever send */ | |
2453 | unsigned int bc_rep_max_size; | |
2454 | }; | |
2455 | ||
2456 | struct ptlrpc_service_thr_conf { | |
2457 | /* threadname should be 8 characters or less - 6 will be added on */ | |
2458 | char *tc_thr_name; | |
2459 | /* threads increasing factor for each CPU */ | |
2460 | unsigned int tc_thr_factor; | |
2461 | /* service threads # to start on each partition while initializing */ | |
2462 | unsigned int tc_nthrs_init; | |
2463 | /* | |
2464 | * low water of threads # upper-limit on each partition while running, | |
2465 | * service availability may be impacted if threads number is lower | |
2466 | * than this value. It can be ZERO if the service doesn't require | |
2467 | * CPU affinity or there is only one partition. | |
2468 | */ | |
2469 | unsigned int tc_nthrs_base; | |
2470 | /* "soft" limit for total threads number */ | |
2471 | unsigned int tc_nthrs_max; | |
2472 | /* user specified threads number, it will be validated due to | |
c56e256d OD |
2473 | * other members of this structure. |
2474 | */ | |
d7e09d03 PT |
2475 | unsigned int tc_nthrs_user; |
2476 | /* set NUMA node affinity for service threads */ | |
2477 | unsigned int tc_cpu_affinity; | |
2478 | /* Tags for lu_context associated with service thread */ | |
2479 | __u32 tc_ctx_tags; | |
2480 | }; | |
2481 | ||
2482 | struct ptlrpc_service_cpt_conf { | |
2483 | struct cfs_cpt_table *cc_cptable; | |
2484 | /* string pattern to describe CPTs for a service */ | |
2485 | char *cc_pattern; | |
2486 | }; | |
2487 | ||
2488 | struct ptlrpc_service_conf { | |
2489 | /* service name */ | |
2490 | char *psc_name; | |
2491 | /* soft watchdog timeout multiplifier to print stuck service traces */ | |
2492 | unsigned int psc_watchdog_factor; | |
2493 | /* buffer information */ | |
2494 | struct ptlrpc_service_buf_conf psc_buf; | |
2495 | /* thread information */ | |
2496 | struct ptlrpc_service_thr_conf psc_thr; | |
2497 | /* CPU partition information */ | |
2498 | struct ptlrpc_service_cpt_conf psc_cpt; | |
2499 | /* function table */ | |
2500 | struct ptlrpc_service_ops psc_ops; | |
2501 | }; | |
2502 | ||
2503 | /* ptlrpc/service.c */ | |
2504 | /** | |
2505 | * Server-side services API. Register/unregister service, request state | |
2506 | * management, service thread management | |
2507 | * | |
2508 | * @{ | |
2509 | */ | |
d7e09d03 PT |
2510 | void ptlrpc_dispatch_difficult_reply(struct ptlrpc_reply_state *rs); |
2511 | void ptlrpc_schedule_difficult_reply(struct ptlrpc_reply_state *rs); | |
10457d4b OD |
2512 | struct ptlrpc_service *ptlrpc_register_service(struct ptlrpc_service_conf *conf, |
2513 | struct kset *parent, | |
2514 | struct dentry *debugfs_entry); | |
d7e09d03 PT |
2515 | |
2516 | int ptlrpc_start_threads(struct ptlrpc_service *svc); | |
2517 | int ptlrpc_unregister_service(struct ptlrpc_service *service); | |
d7e09d03 PT |
2518 | |
2519 | int ptlrpc_hr_init(void); | |
2520 | void ptlrpc_hr_fini(void); | |
2521 | ||
2522 | /** @} */ | |
2523 | ||
2524 | /* ptlrpc/import.c */ | |
2525 | /** | |
2526 | * Import API | |
2527 | * @{ | |
2528 | */ | |
2529 | int ptlrpc_connect_import(struct obd_import *imp); | |
2530 | int ptlrpc_init_import(struct obd_import *imp); | |
2531 | int ptlrpc_disconnect_import(struct obd_import *imp, int noclose); | |
2532 | int ptlrpc_import_recovery_state_machine(struct obd_import *imp); | |
d7e09d03 PT |
2533 | |
2534 | /* ptlrpc/pack_generic.c */ | |
2535 | int ptlrpc_reconnect_import(struct obd_import *imp); | |
2536 | /** @} */ | |
2537 | ||
2538 | /** | |
2539 | * ptlrpc msg buffer and swab interface | |
2540 | * | |
2541 | * @{ | |
2542 | */ | |
2543 | int ptlrpc_buf_need_swab(struct ptlrpc_request *req, const int inout, | |
2544 | int index); | |
2545 | void ptlrpc_buf_set_swabbed(struct ptlrpc_request *req, const int inout, | |
10457d4b | 2546 | int index); |
d7e09d03 PT |
2547 | int ptlrpc_unpack_rep_msg(struct ptlrpc_request *req, int len); |
2548 | int ptlrpc_unpack_req_msg(struct ptlrpc_request *req, int len); | |
2549 | ||
d7e09d03 PT |
2550 | void lustre_init_msg_v2(struct lustre_msg_v2 *msg, int count, __u32 *lens, |
2551 | char **bufs); | |
2552 | int lustre_pack_request(struct ptlrpc_request *, __u32 magic, int count, | |
2553 | __u32 *lens, char **bufs); | |
2554 | int lustre_pack_reply(struct ptlrpc_request *, int count, __u32 *lens, | |
2555 | char **bufs); | |
2556 | int lustre_pack_reply_v2(struct ptlrpc_request *req, int count, | |
2557 | __u32 *lens, char **bufs, int flags); | |
2558 | #define LPRFL_EARLY_REPLY 1 | |
2559 | int lustre_pack_reply_flags(struct ptlrpc_request *, int count, __u32 *lens, | |
2560 | char **bufs, int flags); | |
2561 | int lustre_shrink_msg(struct lustre_msg *msg, int segment, | |
2562 | unsigned int newlen, int move_data); | |
2563 | void lustre_free_reply_state(struct ptlrpc_reply_state *rs); | |
2564 | int __lustre_unpack_msg(struct lustre_msg *m, int len); | |
2565 | int lustre_msg_hdr_size(__u32 magic, int count); | |
2566 | int lustre_msg_size(__u32 magic, int count, __u32 *lengths); | |
2567 | int lustre_msg_size_v2(int count, __u32 *lengths); | |
2568 | int lustre_packed_msg_size(struct lustre_msg *msg); | |
2569 | int lustre_msg_early_size(void); | |
2570 | void *lustre_msg_buf_v2(struct lustre_msg_v2 *m, int n, int min_size); | |
2571 | void *lustre_msg_buf(struct lustre_msg *m, int n, int minlen); | |
2572 | int lustre_msg_buflen(struct lustre_msg *m, int n); | |
d7e09d03 PT |
2573 | int lustre_msg_bufcount(struct lustre_msg *m); |
2574 | char *lustre_msg_string(struct lustre_msg *m, int n, int max_len); | |
2575 | __u32 lustre_msghdr_get_flags(struct lustre_msg *msg); | |
2576 | void lustre_msghdr_set_flags(struct lustre_msg *msg, __u32 flags); | |
2577 | __u32 lustre_msg_get_flags(struct lustre_msg *msg); | |
2578 | void lustre_msg_add_flags(struct lustre_msg *msg, int flags); | |
2579 | void lustre_msg_set_flags(struct lustre_msg *msg, int flags); | |
2580 | void lustre_msg_clear_flags(struct lustre_msg *msg, int flags); | |
2581 | __u32 lustre_msg_get_op_flags(struct lustre_msg *msg); | |
2582 | void lustre_msg_add_op_flags(struct lustre_msg *msg, int flags); | |
d7e09d03 PT |
2583 | struct lustre_handle *lustre_msg_get_handle(struct lustre_msg *msg); |
2584 | __u32 lustre_msg_get_type(struct lustre_msg *msg); | |
d7e09d03 PT |
2585 | void lustre_msg_add_version(struct lustre_msg *msg, int version); |
2586 | __u32 lustre_msg_get_opc(struct lustre_msg *msg); | |
d7e09d03 PT |
2587 | __u64 lustre_msg_get_last_committed(struct lustre_msg *msg); |
2588 | __u64 *lustre_msg_get_versions(struct lustre_msg *msg); | |
2589 | __u64 lustre_msg_get_transno(struct lustre_msg *msg); | |
2590 | __u64 lustre_msg_get_slv(struct lustre_msg *msg); | |
2591 | __u32 lustre_msg_get_limit(struct lustre_msg *msg); | |
2592 | void lustre_msg_set_slv(struct lustre_msg *msg, __u64 slv); | |
2593 | void lustre_msg_set_limit(struct lustre_msg *msg, __u64 limit); | |
2594 | int lustre_msg_get_status(struct lustre_msg *msg); | |
2595 | __u32 lustre_msg_get_conn_cnt(struct lustre_msg *msg); | |
d7e09d03 PT |
2596 | __u32 lustre_msg_get_magic(struct lustre_msg *msg); |
2597 | __u32 lustre_msg_get_timeout(struct lustre_msg *msg); | |
2598 | __u32 lustre_msg_get_service_time(struct lustre_msg *msg); | |
d7e09d03 | 2599 | __u32 lustre_msg_get_cksum(struct lustre_msg *msg); |
d7e09d03 | 2600 | __u32 lustre_msg_calc_cksum(struct lustre_msg *msg); |
1d8cb70c GD |
2601 | void lustre_msg_set_handle(struct lustre_msg *msg, |
2602 | struct lustre_handle *handle); | |
d7e09d03 PT |
2603 | void lustre_msg_set_type(struct lustre_msg *msg, __u32 type); |
2604 | void lustre_msg_set_opc(struct lustre_msg *msg, __u32 opc); | |
d7e09d03 PT |
2605 | void lustre_msg_set_versions(struct lustre_msg *msg, __u64 *versions); |
2606 | void lustre_msg_set_transno(struct lustre_msg *msg, __u64 transno); | |
2607 | void lustre_msg_set_status(struct lustre_msg *msg, __u32 status); | |
2608 | void lustre_msg_set_conn_cnt(struct lustre_msg *msg, __u32 conn_cnt); | |
d7e09d03 PT |
2609 | void ptlrpc_request_set_replen(struct ptlrpc_request *req); |
2610 | void lustre_msg_set_timeout(struct lustre_msg *msg, __u32 timeout); | |
2611 | void lustre_msg_set_service_time(struct lustre_msg *msg, __u32 service_time); | |
2612 | void lustre_msg_set_jobid(struct lustre_msg *msg, char *jobid); | |
2613 | void lustre_msg_set_cksum(struct lustre_msg *msg, __u32 cksum); | |
2614 | ||
2615 | static inline void | |
2616 | lustre_shrink_reply(struct ptlrpc_request *req, int segment, | |
2617 | unsigned int newlen, int move_data) | |
2618 | { | |
2619 | LASSERT(req->rq_reply_state); | |
2620 | LASSERT(req->rq_repmsg); | |
2621 | req->rq_replen = lustre_shrink_msg(req->rq_repmsg, segment, | |
2622 | newlen, move_data); | |
2623 | } | |
2d58de78 LW |
2624 | |
2625 | #ifdef CONFIG_LUSTRE_TRANSLATE_ERRNOS | |
2626 | ||
2627 | static inline int ptlrpc_status_hton(int h) | |
2628 | { | |
2629 | /* | |
2630 | * Positive errnos must be network errnos, such as LUSTRE_EDEADLK, | |
2631 | * ELDLM_LOCK_ABORTED, etc. | |
2632 | */ | |
2633 | if (h < 0) | |
2634 | return -lustre_errno_hton(-h); | |
2635 | else | |
2636 | return h; | |
2637 | } | |
2638 | ||
2639 | static inline int ptlrpc_status_ntoh(int n) | |
2640 | { | |
2641 | /* | |
2642 | * See the comment in ptlrpc_status_hton(). | |
2643 | */ | |
2644 | if (n < 0) | |
2645 | return -lustre_errno_ntoh(-n); | |
2646 | else | |
2647 | return n; | |
2648 | } | |
2649 | ||
2650 | #else | |
2651 | ||
2652 | #define ptlrpc_status_hton(h) (h) | |
2653 | #define ptlrpc_status_ntoh(n) (n) | |
2654 | ||
2655 | #endif | |
d7e09d03 PT |
2656 | /** @} */ |
2657 | ||
2658 | /** Change request phase of \a req to \a new_phase */ | |
2659 | static inline void | |
2660 | ptlrpc_rqphase_move(struct ptlrpc_request *req, enum rq_phase new_phase) | |
2661 | { | |
2662 | if (req->rq_phase == new_phase) | |
2663 | return; | |
2664 | ||
2665 | if (new_phase == RQ_PHASE_UNREGISTERING) { | |
2666 | req->rq_next_phase = req->rq_phase; | |
2667 | if (req->rq_import) | |
2668 | atomic_inc(&req->rq_import->imp_unregistering); | |
2669 | } | |
2670 | ||
2671 | if (req->rq_phase == RQ_PHASE_UNREGISTERING) { | |
2672 | if (req->rq_import) | |
2673 | atomic_dec(&req->rq_import->imp_unregistering); | |
2674 | } | |
2675 | ||
2676 | DEBUG_REQ(D_INFO, req, "move req \"%s\" -> \"%s\"", | |
2677 | ptlrpc_rqphase2str(req), ptlrpc_phase2str(new_phase)); | |
2678 | ||
2679 | req->rq_phase = new_phase; | |
2680 | } | |
2681 | ||
2682 | /** | |
2683 | * Returns true if request \a req got early reply and hard deadline is not met | |
2684 | */ | |
2685 | static inline int | |
2686 | ptlrpc_client_early(struct ptlrpc_request *req) | |
2687 | { | |
2688 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
219e6de6 | 2689 | req->rq_reply_deadline > ktime_get_real_seconds()) |
d7e09d03 PT |
2690 | return 0; |
2691 | return req->rq_early; | |
2692 | } | |
2693 | ||
2694 | /** | |
2695 | * Returns true if we got real reply from server for this request | |
2696 | */ | |
2697 | static inline int | |
2698 | ptlrpc_client_replied(struct ptlrpc_request *req) | |
2699 | { | |
2700 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
219e6de6 | 2701 | req->rq_reply_deadline > ktime_get_real_seconds()) |
d7e09d03 PT |
2702 | return 0; |
2703 | return req->rq_replied; | |
2704 | } | |
2705 | ||
2706 | /** Returns true if request \a req is in process of receiving server reply */ | |
2707 | static inline int | |
2708 | ptlrpc_client_recv(struct ptlrpc_request *req) | |
2709 | { | |
2710 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
219e6de6 | 2711 | req->rq_reply_deadline > ktime_get_real_seconds()) |
d7e09d03 PT |
2712 | return 1; |
2713 | return req->rq_receiving_reply; | |
2714 | } | |
2715 | ||
2716 | static inline int | |
2717 | ptlrpc_client_recv_or_unlink(struct ptlrpc_request *req) | |
2718 | { | |
2719 | int rc; | |
2720 | ||
2721 | spin_lock(&req->rq_lock); | |
2722 | if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) && | |
219e6de6 | 2723 | req->rq_reply_deadline > ktime_get_real_seconds()) { |
d7e09d03 PT |
2724 | spin_unlock(&req->rq_lock); |
2725 | return 1; | |
2726 | } | |
cf378ff7 AL |
2727 | rc = req->rq_receiving_reply; |
2728 | rc = rc || req->rq_req_unlink || req->rq_reply_unlink; | |
d7e09d03 PT |
2729 | spin_unlock(&req->rq_lock); |
2730 | return rc; | |
2731 | } | |
2732 | ||
2733 | static inline void | |
2734 | ptlrpc_client_wake_req(struct ptlrpc_request *req) | |
2735 | { | |
d2a13989 | 2736 | if (!req->rq_set) |
d7e09d03 PT |
2737 | wake_up(&req->rq_reply_waitq); |
2738 | else | |
2739 | wake_up(&req->rq_set->set_waitq); | |
2740 | } | |
2741 | ||
2742 | static inline void | |
2743 | ptlrpc_rs_addref(struct ptlrpc_reply_state *rs) | |
2744 | { | |
2745 | LASSERT(atomic_read(&rs->rs_refcount) > 0); | |
2746 | atomic_inc(&rs->rs_refcount); | |
2747 | } | |
2748 | ||
2749 | static inline void | |
2750 | ptlrpc_rs_decref(struct ptlrpc_reply_state *rs) | |
2751 | { | |
2752 | LASSERT(atomic_read(&rs->rs_refcount) > 0); | |
2753 | if (atomic_dec_and_test(&rs->rs_refcount)) | |
2754 | lustre_free_reply_state(rs); | |
2755 | } | |
2756 | ||
2757 | /* Should only be called once per req */ | |
2758 | static inline void ptlrpc_req_drop_rs(struct ptlrpc_request *req) | |
2759 | { | |
d2a13989 | 2760 | if (!req->rq_reply_state) |
d7e09d03 PT |
2761 | return; /* shouldn't occur */ |
2762 | ptlrpc_rs_decref(req->rq_reply_state); | |
2763 | req->rq_reply_state = NULL; | |
2764 | req->rq_repmsg = NULL; | |
2765 | } | |
2766 | ||
2767 | static inline __u32 lustre_request_magic(struct ptlrpc_request *req) | |
2768 | { | |
2769 | return lustre_msg_get_magic(req->rq_reqmsg); | |
2770 | } | |
2771 | ||
2772 | static inline int ptlrpc_req_get_repsize(struct ptlrpc_request *req) | |
2773 | { | |
2774 | switch (req->rq_reqmsg->lm_magic) { | |
2775 | case LUSTRE_MSG_MAGIC_V2: | |
2776 | return req->rq_reqmsg->lm_repsize; | |
2777 | default: | |
2778 | LASSERTF(0, "incorrect message magic: %08x\n", | |
2779 | req->rq_reqmsg->lm_magic); | |
2780 | return -EFAULT; | |
2781 | } | |
2782 | } | |
2783 | ||
2784 | static inline int ptlrpc_send_limit_expired(struct ptlrpc_request *req) | |
2785 | { | |
2786 | if (req->rq_delay_limit != 0 && | |
699503bc GKH |
2787 | time_before(cfs_time_add(req->rq_queued_time, |
2788 | cfs_time_seconds(req->rq_delay_limit)), | |
2789 | cfs_time_current())) { | |
d7e09d03 PT |
2790 | return 1; |
2791 | } | |
2792 | return 0; | |
2793 | } | |
2794 | ||
2795 | static inline int ptlrpc_no_resend(struct ptlrpc_request *req) | |
2796 | { | |
2797 | if (!req->rq_no_resend && ptlrpc_send_limit_expired(req)) { | |
2798 | spin_lock(&req->rq_lock); | |
2799 | req->rq_no_resend = 1; | |
2800 | spin_unlock(&req->rq_lock); | |
2801 | } | |
2802 | return req->rq_no_resend; | |
2803 | } | |
2804 | ||
2805 | static inline int | |
2806 | ptlrpc_server_get_timeout(struct ptlrpc_service_part *svcpt) | |
2807 | { | |
2808 | int at = AT_OFF ? 0 : at_get(&svcpt->scp_at_estimate); | |
2809 | ||
2810 | return svcpt->scp_service->srv_watchdog_factor * | |
2811 | max_t(int, at, obd_timeout); | |
2812 | } | |
2813 | ||
2814 | static inline struct ptlrpc_service * | |
2815 | ptlrpc_req2svc(struct ptlrpc_request *req) | |
2816 | { | |
d7e09d03 PT |
2817 | return req->rq_rqbd->rqbd_svcpt->scp_service; |
2818 | } | |
2819 | ||
2820 | /* ldlm/ldlm_lib.c */ | |
2821 | /** | |
2822 | * Target client logic | |
2823 | * @{ | |
2824 | */ | |
2825 | int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg); | |
2826 | int client_obd_cleanup(struct obd_device *obddev); | |
2827 | int client_connect_import(const struct lu_env *env, | |
2828 | struct obd_export **exp, struct obd_device *obd, | |
2829 | struct obd_uuid *cluuid, struct obd_connect_data *, | |
2830 | void *localdata); | |
2831 | int client_disconnect_export(struct obd_export *exp); | |
2832 | int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid, | |
2833 | int priority); | |
2834 | int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid); | |
2835 | int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, | |
2836 | struct obd_uuid *uuid); | |
2837 | int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid); | |
2838 | void client_destroy_import(struct obd_import *imp); | |
2839 | /** @} */ | |
2840 | ||
d7e09d03 PT |
2841 | /* ptlrpc/pinger.c */ |
2842 | /** | |
2843 | * Pinger API (client side only) | |
2844 | * @{ | |
2845 | */ | |
2846 | enum timeout_event { | |
2847 | TIMEOUT_GRANT = 1 | |
2848 | }; | |
c9f6bb96 | 2849 | |
d7e09d03 PT |
2850 | struct timeout_item; |
2851 | typedef int (*timeout_cb_t)(struct timeout_item *, void *); | |
2852 | int ptlrpc_pinger_add_import(struct obd_import *imp); | |
2853 | int ptlrpc_pinger_del_import(struct obd_import *imp); | |
2854 | int ptlrpc_add_timeout_client(int time, enum timeout_event event, | |
2855 | timeout_cb_t cb, void *data, | |
2856 | struct list_head *obd_list); | |
2857 | int ptlrpc_del_timeout_client(struct list_head *obd_list, | |
2858 | enum timeout_event event); | |
aff9d8e8 | 2859 | struct ptlrpc_request *ptlrpc_prep_ping(struct obd_import *imp); |
d7e09d03 | 2860 | int ptlrpc_obd_ping(struct obd_device *obd); |
d7e09d03 PT |
2861 | void ptlrpc_pinger_ir_up(void); |
2862 | void ptlrpc_pinger_ir_down(void); | |
2863 | /** @} */ | |
2864 | int ptlrpc_pinger_suppress_pings(void); | |
2865 | ||
d7e09d03 PT |
2866 | /* ptlrpc/ptlrpcd.c */ |
2867 | void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force); | |
2868 | void ptlrpcd_free(struct ptlrpcd_ctl *pc); | |
2869 | void ptlrpcd_wake(struct ptlrpc_request *req); | |
c5c4c6fa | 2870 | void ptlrpcd_add_req(struct ptlrpc_request *req); |
d7e09d03 PT |
2871 | int ptlrpcd_addref(void); |
2872 | void ptlrpcd_decref(void); | |
2873 | ||
2874 | /* ptlrpc/lproc_ptlrpc.c */ | |
2875 | /** | |
2876 | * procfs output related functions | |
2877 | * @{ | |
2878 | */ | |
9c234f6c | 2879 | const char *ll_opcode2str(__u32 opcode); |
d7e09d03 PT |
2880 | void ptlrpc_lprocfs_register_obd(struct obd_device *obd); |
2881 | void ptlrpc_lprocfs_unregister_obd(struct obd_device *obd); | |
2882 | void ptlrpc_lprocfs_brw(struct ptlrpc_request *req, int bytes); | |
d7e09d03 PT |
2883 | /** @} */ |
2884 | ||
d7e09d03 PT |
2885 | /* ptlrpc/llog_client.c */ |
2886 | extern struct llog_operations llog_client_ops; | |
d7e09d03 PT |
2887 | /** @} net */ |
2888 | ||
2889 | #endif | |
2890 | /** @} PtlRPC */ |