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) 2003, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | */ | |
36 | ||
37 | #ifndef __LNET_TYPES_H__ | |
38 | #define __LNET_TYPES_H__ | |
39 | ||
db18b8e9 JS |
40 | #include <linux/types.h> |
41 | ||
d7e09d03 PT |
42 | /** \addtogroup lnet |
43 | * @{ */ | |
44 | ||
d7e09d03 PT |
45 | /** \addtogroup lnet_addr |
46 | * @{ */ | |
47 | ||
48 | /** Portal reserved for LNet's own use. | |
49 | * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments. | |
50 | */ | |
51 | #define LNET_RESERVED_PORTAL 0 | |
52 | ||
53 | /** | |
54 | * Address of an end-point in an LNet network. | |
55 | * | |
56 | * A node can have multiple end-points and hence multiple addresses. | |
57 | * An LNet network can be a simple network (e.g. tcp0) or a network of | |
58 | * LNet networks connected by LNet routers. Therefore an end-point address | |
59 | * has two parts: network ID, and address within a network. | |
60 | * | |
61 | * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID. | |
62 | */ | |
63 | typedef __u64 lnet_nid_t; | |
64 | /** | |
65 | * ID of a process in a node. Shortened as PID to distinguish from | |
66 | * lnet_process_id_t, the global process ID. | |
67 | */ | |
68 | typedef __u32 lnet_pid_t; | |
69 | ||
70 | /** wildcard NID that matches any end-point address */ | |
71 | #define LNET_NID_ANY ((lnet_nid_t) -1) | |
72 | /** wildcard PID that matches any lnet_pid_t */ | |
73 | #define LNET_PID_ANY ((lnet_pid_t) -1) | |
74 | ||
75 | #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */ | |
76 | #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */ | |
db18b8e9 | 77 | #define LNET_PID_LUSTRE 12345 |
d7e09d03 PT |
78 | |
79 | #define LNET_TIME_FOREVER (-1) | |
80 | ||
bbf00c3d JS |
81 | /* how an LNET NID encodes net:address */ |
82 | /** extract the address part of an lnet_nid_t */ | |
83 | ||
84 | static inline __u32 LNET_NIDADDR(lnet_nid_t nid) | |
85 | { | |
86 | return nid & 0xffffffff; | |
87 | } | |
88 | ||
89 | static inline __u32 LNET_NIDNET(lnet_nid_t nid) | |
90 | { | |
91 | return (nid >> 32) & 0xffffffff; | |
92 | } | |
93 | ||
94 | static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr) | |
95 | { | |
96 | return (((__u64)net) << 32) | addr; | |
97 | } | |
98 | ||
99 | static inline __u32 LNET_NETNUM(__u32 net) | |
100 | { | |
101 | return net & 0xffff; | |
102 | } | |
103 | ||
104 | static inline __u32 LNET_NETTYP(__u32 net) | |
105 | { | |
106 | return (net >> 16) & 0xffff; | |
107 | } | |
108 | ||
109 | static inline __u32 LNET_MKNET(__u32 type, __u32 num) | |
110 | { | |
111 | return (type << 16) | num; | |
112 | } | |
113 | ||
db18b8e9 JS |
114 | #define WIRE_ATTR __packed |
115 | ||
116 | /* Packed version of lnet_process_id_t to transfer via network */ | |
117 | typedef struct { | |
118 | /* node id / process id */ | |
119 | lnet_nid_t nid; | |
120 | lnet_pid_t pid; | |
121 | } WIRE_ATTR lnet_process_id_packed_t; | |
122 | ||
123 | /* The wire handle's interface cookie only matches one network interface in | |
124 | * one epoch (i.e. new cookie when the interface restarts or the node | |
125 | * reboots). The object cookie only matches one object on that interface | |
126 | * during that object's lifetime (i.e. no cookie re-use). */ | |
127 | typedef struct { | |
128 | __u64 wh_interface_cookie; | |
129 | __u64 wh_object_cookie; | |
130 | } WIRE_ATTR lnet_handle_wire_t; | |
131 | ||
132 | typedef enum { | |
133 | LNET_MSG_ACK = 0, | |
134 | LNET_MSG_PUT, | |
135 | LNET_MSG_GET, | |
136 | LNET_MSG_REPLY, | |
137 | LNET_MSG_HELLO, | |
138 | } lnet_msg_type_t; | |
139 | ||
140 | /* The variant fields of the portals message header are aligned on an 8 | |
141 | * byte boundary in the message header. Note that all types used in these | |
142 | * wire structs MUST be fixed size and the smaller types are placed at the | |
143 | * end. */ | |
144 | typedef struct lnet_ack { | |
145 | lnet_handle_wire_t dst_wmd; | |
146 | __u64 match_bits; | |
147 | __u32 mlength; | |
148 | } WIRE_ATTR lnet_ack_t; | |
149 | ||
150 | typedef struct lnet_put { | |
151 | lnet_handle_wire_t ack_wmd; | |
152 | __u64 match_bits; | |
153 | __u64 hdr_data; | |
154 | __u32 ptl_index; | |
155 | __u32 offset; | |
156 | } WIRE_ATTR lnet_put_t; | |
157 | ||
158 | typedef struct lnet_get { | |
159 | lnet_handle_wire_t return_wmd; | |
160 | __u64 match_bits; | |
161 | __u32 ptl_index; | |
162 | __u32 src_offset; | |
163 | __u32 sink_length; | |
164 | } WIRE_ATTR lnet_get_t; | |
165 | ||
166 | typedef struct lnet_reply { | |
167 | lnet_handle_wire_t dst_wmd; | |
168 | } WIRE_ATTR lnet_reply_t; | |
169 | ||
170 | typedef struct lnet_hello { | |
171 | __u64 incarnation; | |
172 | __u32 type; | |
173 | } WIRE_ATTR lnet_hello_t; | |
174 | ||
175 | typedef struct { | |
176 | lnet_nid_t dest_nid; | |
177 | lnet_nid_t src_nid; | |
178 | lnet_pid_t dest_pid; | |
179 | lnet_pid_t src_pid; | |
180 | __u32 type; /* lnet_msg_type_t */ | |
181 | __u32 payload_length; /* payload data to follow */ | |
182 | /*<------__u64 aligned------->*/ | |
183 | union { | |
184 | lnet_ack_t ack; | |
185 | lnet_put_t put; | |
186 | lnet_get_t get; | |
187 | lnet_reply_t reply; | |
188 | lnet_hello_t hello; | |
189 | } msg; | |
190 | } WIRE_ATTR lnet_hdr_t; | |
191 | ||
192 | /* A HELLO message contains a magic number and protocol version | |
193 | * code in the header's dest_nid, the peer's NID in the src_nid, and | |
194 | * LNET_MSG_HELLO in the type field. All other common fields are zero | |
195 | * (including payload_size; i.e. no payload). | |
196 | * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is | |
197 | * running the same protocol and to find out its NID. These LNDs should | |
198 | * exchange HELLO messages when a connection is first established. Individual | |
199 | * LNDs can put whatever else they fancy in lnet_hdr_t::msg. | |
200 | */ | |
201 | typedef struct { | |
202 | __u32 magic; /* LNET_PROTO_TCP_MAGIC */ | |
203 | __u16 version_major; /* increment on incompatible change */ | |
204 | __u16 version_minor; /* increment on compatible change */ | |
205 | } WIRE_ATTR lnet_magicversion_t; | |
206 | ||
207 | /* PROTO MAGIC for LNDs */ | |
208 | #define LNET_PROTO_IB_MAGIC 0x0be91b91 | |
209 | #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */ | |
210 | #define LNET_PROTO_TCP_MAGIC 0xeebc0ded | |
211 | #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100 | |
212 | #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */ | |
213 | ||
214 | /* Placeholder for a future "unified" protocol across all LNDs */ | |
215 | /* Current LNDs that receive a request with this magic will respond with a | |
216 | * "stub" reply using their current protocol */ | |
217 | #define LNET_PROTO_MAGIC 0x45726963 /* ! */ | |
218 | ||
219 | #define LNET_PROTO_TCP_VERSION_MAJOR 1 | |
220 | #define LNET_PROTO_TCP_VERSION_MINOR 0 | |
221 | ||
222 | /* Acceptor connection request */ | |
223 | typedef struct { | |
224 | __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */ | |
225 | __u32 acr_version; /* protocol version */ | |
226 | __u64 acr_nid; /* target NID */ | |
227 | } WIRE_ATTR lnet_acceptor_connreq_t; | |
228 | ||
229 | #define LNET_PROTO_ACCEPTOR_VERSION 1 | |
230 | ||
231 | typedef struct { | |
232 | lnet_nid_t ns_nid; | |
233 | __u32 ns_status; | |
234 | __u32 ns_unused; | |
235 | } WIRE_ATTR lnet_ni_status_t; | |
236 | ||
237 | typedef struct { | |
238 | __u32 pi_magic; | |
239 | __u32 pi_features; | |
240 | lnet_pid_t pi_pid; | |
241 | __u32 pi_nnis; | |
242 | lnet_ni_status_t pi_ni[0]; | |
243 | } WIRE_ATTR lnet_ping_info_t; | |
244 | ||
245 | typedef struct lnet_counters { | |
246 | __u32 msgs_alloc; | |
247 | __u32 msgs_max; | |
248 | __u32 errors; | |
249 | __u32 send_count; | |
250 | __u32 recv_count; | |
251 | __u32 route_count; | |
252 | __u32 drop_count; | |
253 | __u64 send_length; | |
254 | __u64 recv_length; | |
255 | __u64 route_length; | |
256 | __u64 drop_length; | |
257 | } WIRE_ATTR lnet_counters_t; | |
258 | ||
259 | #define LNET_NI_STATUS_UP 0x15aac0de | |
260 | #define LNET_NI_STATUS_DOWN 0xdeadface | |
261 | #define LNET_NI_STATUS_INVALID 0x00000000 | |
262 | ||
263 | #define LNET_MAX_INTERFACES 16 | |
264 | ||
265 | /* | |
d7e09d03 PT |
266 | * Objects maintained by the LNet are accessed through handles. Handle types |
267 | * have names of the form lnet_handle_xx_t, where xx is one of the two letter | |
268 | * object type codes ('eq' for event queue, 'md' for memory descriptor, and | |
269 | * 'me' for match entry). | |
270 | * Each type of object is given a unique handle type to enhance type checking. | |
271 | * The type lnet_handle_any_t can be used when a generic handle is needed. | |
272 | * Every handle value can be converted into a value of type lnet_handle_any_t | |
273 | * without loss of information. | |
274 | */ | |
275 | typedef struct { | |
276 | __u64 cookie; | |
277 | } lnet_handle_any_t; | |
278 | ||
279 | typedef lnet_handle_any_t lnet_handle_eq_t; | |
280 | typedef lnet_handle_any_t lnet_handle_md_t; | |
281 | typedef lnet_handle_any_t lnet_handle_me_t; | |
282 | ||
283 | #define LNET_WIRE_HANDLE_COOKIE_NONE (-1) | |
284 | ||
285 | /** | |
286 | * Invalidate handle \a h. | |
287 | */ | |
288 | static inline void LNetInvalidateHandle(lnet_handle_any_t *h) | |
289 | { | |
290 | h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE; | |
291 | } | |
292 | ||
293 | /** | |
294 | * Compare handles \a h1 and \a h2. | |
295 | * | |
296 | * \return 1 if handles are equal, 0 if otherwise. | |
297 | */ | |
568414f1 | 298 | static inline int LNetHandleIsEqual(lnet_handle_any_t h1, lnet_handle_any_t h2) |
d7e09d03 | 299 | { |
e4422943 | 300 | return h1.cookie == h2.cookie; |
d7e09d03 PT |
301 | } |
302 | ||
303 | /** | |
304 | * Check whether handle \a h is invalid. | |
305 | * | |
306 | * \return 1 if handle is invalid, 0 if valid. | |
307 | */ | |
308 | static inline int LNetHandleIsInvalid(lnet_handle_any_t h) | |
309 | { | |
e4422943 | 310 | return LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie; |
d7e09d03 PT |
311 | } |
312 | ||
313 | /** | |
314 | * Global process ID. | |
315 | */ | |
316 | typedef struct { | |
317 | /** node id */ | |
318 | lnet_nid_t nid; | |
319 | /** process id */ | |
320 | lnet_pid_t pid; | |
321 | } lnet_process_id_t; | |
322 | /** @} lnet_addr */ | |
323 | ||
324 | /** \addtogroup lnet_me | |
325 | * @{ */ | |
326 | ||
327 | /** | |
328 | * Specifies whether the match entry or memory descriptor should be unlinked | |
329 | * automatically (LNET_UNLINK) or not (LNET_RETAIN). | |
330 | */ | |
331 | typedef enum { | |
332 | LNET_RETAIN = 0, | |
333 | LNET_UNLINK | |
334 | } lnet_unlink_t; | |
335 | ||
336 | /** | |
337 | * Values of the type lnet_ins_pos_t are used to control where a new match | |
338 | * entry is inserted. The value LNET_INS_BEFORE is used to insert the new | |
339 | * entry before the current entry or before the head of the list. The value | |
340 | * LNET_INS_AFTER is used to insert the new entry after the current entry | |
341 | * or after the last item in the list. | |
342 | */ | |
343 | typedef enum { | |
344 | /** insert ME before current position or head of the list */ | |
345 | LNET_INS_BEFORE, | |
346 | /** insert ME after current position or tail of the list */ | |
347 | LNET_INS_AFTER, | |
348 | /** attach ME at tail of local CPU partition ME list */ | |
349 | LNET_INS_LOCAL | |
350 | } lnet_ins_pos_t; | |
351 | ||
352 | /** @} lnet_me */ | |
353 | ||
354 | /** \addtogroup lnet_md | |
355 | * @{ */ | |
356 | ||
357 | /** | |
358 | * Defines the visible parts of a memory descriptor. Values of this type | |
359 | * are used to initialize memory descriptors. | |
360 | */ | |
361 | typedef struct { | |
362 | /** | |
363 | * Specify the memory region associated with the memory descriptor. | |
364 | * If the options field has: | |
365 | * - LNET_MD_KIOV bit set: The start field points to the starting | |
366 | * address of an array of lnet_kiov_t and the length field specifies | |
367 | * the number of entries in the array. The length can't be bigger | |
368 | * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based | |
d766b4b5 | 369 | * fragments that are not necessarily mapped in virtual memory. |
d7e09d03 PT |
370 | * - LNET_MD_IOVEC bit set: The start field points to the starting |
371 | * address of an array of struct iovec and the length field specifies | |
372 | * the number of entries in the array. The length can't be bigger | |
373 | * than LNET_MAX_IOV. The struct iovec is used to describe fragments | |
374 | * that have virtual addresses. | |
375 | * - Otherwise: The memory region is contiguous. The start field | |
376 | * specifies the starting address for the memory region and the | |
377 | * length field specifies its length. | |
378 | * | |
379 | * When the memory region is fragmented, all fragments but the first | |
380 | * one must start on page boundary, and all but the last must end on | |
381 | * page boundary. | |
382 | */ | |
383 | void *start; | |
384 | unsigned int length; | |
385 | /** | |
386 | * Specifies the maximum number of operations that can be performed | |
387 | * on the memory descriptor. An operation is any action that could | |
388 | * possibly generate an event. In the usual case, the threshold value | |
389 | * is decremented for each operation on the MD. When the threshold | |
390 | * drops to zero, the MD becomes inactive and does not respond to | |
391 | * operations. A threshold value of LNET_MD_THRESH_INF indicates that | |
392 | * there is no bound on the number of operations that may be applied | |
393 | * to a MD. | |
394 | */ | |
395 | int threshold; | |
396 | /** | |
397 | * Specifies the largest incoming request that the memory descriptor | |
398 | * should respond to. When the unused portion of a MD (length - | |
399 | * local offset) falls below this value, the MD becomes inactive and | |
400 | * does not respond to further operations. This value is only used | |
401 | * if the LNET_MD_MAX_SIZE option is set. | |
402 | */ | |
403 | int max_size; | |
404 | /** | |
405 | * Specifies the behavior of the memory descriptor. A bitwise OR | |
406 | * of the following values can be used: | |
407 | * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD. | |
408 | * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD. | |
409 | * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory | |
410 | * region is provided by the incoming request. By default, the | |
411 | * offset is maintained locally. When maintained locally, the | |
412 | * offset is incremented by the length of the request so that | |
413 | * the next operation (PUT or GET) will access the next part of | |
414 | * the memory region. Note that only one offset variable exists | |
415 | * per memory descriptor. If both PUT and GET operations are | |
416 | * performed on a memory descriptor, the offset is updated each time. | |
417 | * - LNET_MD_TRUNCATE: The length provided in the incoming request can | |
418 | * be reduced to match the memory available in the region (determined | |
419 | * by subtracting the offset from the length of the memory region). | |
420 | * By default, if the length in the incoming operation is greater | |
421 | * than the amount of memory available, the operation is rejected. | |
422 | * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for | |
423 | * incoming PUT operations, even if requested. By default, | |
424 | * acknowledgments are sent for PUT operations that request an | |
425 | * acknowledgment. Acknowledgments are never sent for GET operations. | |
426 | * The data sent in the REPLY serves as an implicit acknowledgment. | |
427 | * - LNET_MD_KIOV: The start and length fields specify an array of | |
428 | * lnet_kiov_t. | |
429 | * - LNET_MD_IOVEC: The start and length fields specify an array of | |
430 | * struct iovec. | |
431 | * - LNET_MD_MAX_SIZE: The max_size field is valid. | |
432 | * | |
433 | * Note: | |
434 | * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather | |
435 | * capability for memory descriptors. They can't be both set. | |
436 | * - When LNET_MD_MAX_SIZE is set, the total length of the memory | |
437 | * region (i.e. sum of all fragment lengths) must not be less than | |
438 | * \a max_size. | |
439 | */ | |
440 | unsigned int options; | |
441 | /** | |
442 | * A user-specified value that is associated with the memory | |
443 | * descriptor. The value does not need to be a pointer, but must fit | |
444 | * in the space used by a pointer. This value is recorded in events | |
445 | * associated with operations on this MD. | |
446 | */ | |
447 | void *user_ptr; | |
448 | /** | |
449 | * A handle for the event queue used to log the operations performed on | |
450 | * the memory region. If this argument is a NULL handle (i.e. nullified | |
451 | * by LNetInvalidateHandle()), operations performed on this memory | |
452 | * descriptor are not logged. | |
453 | */ | |
454 | lnet_handle_eq_t eq_handle; | |
455 | } lnet_md_t; | |
456 | ||
457 | /* Max Transfer Unit (minimum supported everywhere). | |
458 | * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks) | |
459 | * these limits are system wide and not interface-local. */ | |
460 | #define LNET_MTU_BITS 20 | |
461 | #define LNET_MTU (1 << LNET_MTU_BITS) | |
462 | ||
463 | /** limit on the number of fragments in discontiguous MDs */ | |
464 | #define LNET_MAX_IOV 256 | |
465 | ||
d7e09d03 PT |
466 | /** |
467 | * Options for the MD structure. See lnet_md_t::options. | |
468 | */ | |
469 | #define LNET_MD_OP_PUT (1 << 0) | |
470 | /** See lnet_md_t::options. */ | |
471 | #define LNET_MD_OP_GET (1 << 1) | |
472 | /** See lnet_md_t::options. */ | |
473 | #define LNET_MD_MANAGE_REMOTE (1 << 2) | |
474 | /* unused (1 << 3) */ | |
475 | /** See lnet_md_t::options. */ | |
476 | #define LNET_MD_TRUNCATE (1 << 4) | |
477 | /** See lnet_md_t::options. */ | |
478 | #define LNET_MD_ACK_DISABLE (1 << 5) | |
479 | /** See lnet_md_t::options. */ | |
480 | #define LNET_MD_IOVEC (1 << 6) | |
481 | /** See lnet_md_t::options. */ | |
482 | #define LNET_MD_MAX_SIZE (1 << 7) | |
483 | /** See lnet_md_t::options. */ | |
484 | #define LNET_MD_KIOV (1 << 8) | |
485 | ||
486 | /* For compatibility with Cray Portals */ | |
487 | #define LNET_MD_PHYS 0 | |
488 | ||
489 | /** Infinite threshold on MD operations. See lnet_md_t::threshold */ | |
490 | #define LNET_MD_THRESH_INF (-1) | |
491 | ||
492 | /* NB lustre portals uses struct iovec internally! */ | |
493 | typedef struct iovec lnet_md_iovec_t; | |
494 | ||
495 | /** | |
496 | * A page-based fragment of a MD. | |
497 | */ | |
498 | typedef struct { | |
499 | /** Pointer to the page where the fragment resides */ | |
500 | struct page *kiov_page; | |
501 | /** Length in bytes of the fragment */ | |
502 | unsigned int kiov_len; | |
503 | /** | |
504 | * Starting offset of the fragment within the page. Note that the | |
505 | * end of the fragment must not pass the end of the page; i.e., | |
506 | * kiov_len + kiov_offset <= PAGE_CACHE_SIZE. | |
507 | */ | |
508 | unsigned int kiov_offset; | |
509 | } lnet_kiov_t; | |
510 | /** @} lnet_md */ | |
511 | ||
512 | /** \addtogroup lnet_eq | |
513 | * @{ */ | |
514 | ||
515 | /** | |
516 | * Six types of events can be logged in an event queue. | |
517 | */ | |
518 | typedef enum { | |
519 | /** An incoming GET operation has completed on the MD. */ | |
520 | LNET_EVENT_GET = 1, | |
521 | /** | |
522 | * An incoming PUT operation has completed on the MD. The | |
523 | * underlying layers will not alter the memory (on behalf of this | |
524 | * operation) once this event has been logged. | |
525 | */ | |
526 | LNET_EVENT_PUT, | |
527 | /** | |
528 | * A REPLY operation has completed. This event is logged after the | |
529 | * data (if any) from the REPLY has been written into the MD. | |
530 | */ | |
531 | LNET_EVENT_REPLY, | |
532 | /** An acknowledgment has been received. */ | |
533 | LNET_EVENT_ACK, | |
534 | /** | |
535 | * An outgoing send (PUT or GET) operation has completed. This event | |
536 | * is logged after the entire buffer has been sent and it is safe for | |
537 | * the caller to reuse the buffer. | |
538 | * | |
539 | * Note: | |
540 | * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can | |
541 | * happen even when the message has not yet been put out on wire. | |
542 | * - It's unsafe to assume that in an outgoing GET operation | |
543 | * the LNET_EVENT_SEND event would happen before the | |
544 | * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and | |
545 | * LNET_EVENT_ACK events in an outgoing PUT operation. | |
546 | */ | |
547 | LNET_EVENT_SEND, | |
548 | /** | |
549 | * A MD has been unlinked. Note that LNetMDUnlink() does not | |
550 | * necessarily trigger an LNET_EVENT_UNLINK event. | |
551 | * \see LNetMDUnlink | |
552 | */ | |
553 | LNET_EVENT_UNLINK, | |
554 | } lnet_event_kind_t; | |
555 | ||
556 | #define LNET_SEQ_BASETYPE long | |
557 | typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t; | |
e4422943 | 558 | #define LNET_SEQ_GT(a, b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0) |
d7e09d03 | 559 | |
d7e09d03 PT |
560 | /** |
561 | * Information about an event on a MD. | |
562 | */ | |
563 | typedef struct { | |
564 | /** The identifier (nid, pid) of the target. */ | |
565 | lnet_process_id_t target; | |
566 | /** The identifier (nid, pid) of the initiator. */ | |
567 | lnet_process_id_t initiator; | |
568 | /** | |
569 | * The NID of the immediate sender. If the request has been forwarded | |
570 | * by routers, this is the NID of the last hop; otherwise it's the | |
571 | * same as the initiator. | |
572 | */ | |
573 | lnet_nid_t sender; | |
574 | /** Indicates the type of the event. */ | |
575 | lnet_event_kind_t type; | |
576 | /** The portal table index specified in the request */ | |
577 | unsigned int pt_index; | |
578 | /** A copy of the match bits specified in the request. */ | |
579 | __u64 match_bits; | |
580 | /** The length (in bytes) specified in the request. */ | |
581 | unsigned int rlength; | |
582 | /** | |
583 | * The length (in bytes) of the data that was manipulated by the | |
584 | * operation. For truncated operations, the manipulated length will be | |
585 | * the number of bytes specified by the MD (possibly with an offset, | |
586 | * see lnet_md_t). For all other operations, the manipulated length | |
587 | * will be the length of the requested operation, i.e. rlength. | |
588 | */ | |
589 | unsigned int mlength; | |
590 | /** | |
591 | * The handle to the MD associated with the event. The handle may be | |
592 | * invalid if the MD has been unlinked. | |
593 | */ | |
594 | lnet_handle_md_t md_handle; | |
595 | /** | |
596 | * A snapshot of the state of the MD immediately after the event has | |
597 | * been processed. In particular, the threshold field in md will | |
598 | * reflect the value of the threshold after the operation occurred. | |
599 | */ | |
600 | lnet_md_t md; | |
601 | /** | |
602 | * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT. | |
603 | * \see LNetPut | |
604 | */ | |
605 | __u64 hdr_data; | |
606 | /** | |
607 | * Indicates the completion status of the operation. It's 0 for | |
608 | * successful operations, otherwise it's an error code. | |
609 | */ | |
610 | int status; | |
611 | /** | |
612 | * Indicates whether the MD has been unlinked. Note that: | |
613 | * - An event with unlinked set is the last event on the MD. | |
614 | * - This field is also set for an explicit LNET_EVENT_UNLINK event. | |
615 | * \see LNetMDUnlink | |
616 | */ | |
617 | int unlinked; | |
618 | /** | |
619 | * The displacement (in bytes) into the memory region that the | |
620 | * operation used. The offset can be determined by the operation for | |
621 | * a remote managed MD or by the local MD. | |
622 | * \see lnet_md_t::options | |
623 | */ | |
624 | unsigned int offset; | |
625 | /** | |
626 | * The sequence number for this event. Sequence numbers are unique | |
627 | * to each event. | |
628 | */ | |
629 | volatile lnet_seq_t sequence; | |
630 | } lnet_event_t; | |
d7e09d03 PT |
631 | |
632 | /** | |
633 | * Event queue handler function type. | |
634 | * | |
635 | * The EQ handler runs for each event that is deposited into the EQ. The | |
636 | * handler is supplied with a pointer to the event that triggered the | |
637 | * handler invocation. | |
638 | * | |
639 | * The handler must not block, must be reentrant, and must not call any LNet | |
640 | * API functions. It should return as quickly as possible. | |
641 | */ | |
642 | typedef void (*lnet_eq_handler_t)(lnet_event_t *event); | |
643 | #define LNET_EQ_HANDLER_NONE NULL | |
644 | /** @} lnet_eq */ | |
645 | ||
646 | /** \addtogroup lnet_data | |
647 | * @{ */ | |
648 | ||
649 | /** | |
650 | * Specify whether an acknowledgment should be sent by target when the PUT | |
651 | * operation completes (i.e., when the data has been written to a MD of the | |
652 | * target process). | |
653 | * | |
654 | * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which | |
655 | * acknowledgments can be disabled for a MD. | |
656 | */ | |
657 | typedef enum { | |
658 | /** Request an acknowledgment */ | |
659 | LNET_ACK_REQ, | |
660 | /** Request that no acknowledgment should be generated. */ | |
661 | LNET_NOACK_REQ | |
662 | } lnet_ack_req_t; | |
663 | /** @} lnet_data */ | |
664 | ||
665 | /** @} lnet */ | |
666 | #endif |