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 | |
4f3ca893 | 18 | * http://www.gnu.org/licenses/gpl-2.0.html |
d7e09d03 PT |
19 | * |
20 | * GPL HEADER END | |
21 | */ | |
22 | /* | |
23 | * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. | |
24 | * Use is subject to license terms. | |
25 | * | |
4f3ca893 | 26 | * Copyright (c) 2012 - 2015, Intel Corporation. |
d7e09d03 PT |
27 | */ |
28 | /* | |
29 | * This file is part of Lustre, http://www.lustre.org/ | |
4f3ca893 | 30 | * Lustre is a trademark of Seagate, Inc. |
d7e09d03 PT |
31 | */ |
32 | ||
33 | #ifndef __LNET_TYPES_H__ | |
34 | #define __LNET_TYPES_H__ | |
35 | ||
db18b8e9 JS |
36 | #include <linux/types.h> |
37 | ||
d7e09d03 | 38 | /** \addtogroup lnet |
4420cfd3 JS |
39 | * @{ |
40 | */ | |
d7e09d03 | 41 | |
5b0e50b9 JS |
42 | #define LNET_VERSION "0.6.0" |
43 | ||
d7e09d03 | 44 | /** \addtogroup lnet_addr |
4420cfd3 JS |
45 | * @{ |
46 | */ | |
d7e09d03 PT |
47 | |
48 | /** Portal reserved for LNet's own use. | |
49 | * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments. | |
50 | */ | |
188acc61 | 51 | #define LNET_RESERVED_PORTAL 0 |
d7e09d03 PT |
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 */ | |
188acc61 | 71 | #define LNET_NID_ANY ((lnet_nid_t) -1) |
d7e09d03 | 72 | /** wildcard PID that matches any lnet_pid_t */ |
188acc61 | 73 | #define LNET_PID_ANY ((lnet_pid_t) -1) |
d7e09d03 PT |
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 | 78 | |
188acc61 | 79 | #define LNET_TIME_FOREVER (-1) |
d7e09d03 | 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 | ||
4420cfd3 JS |
123 | /* |
124 | * The wire handle's interface cookie only matches one network interface in | |
db18b8e9 JS |
125 | * one epoch (i.e. new cookie when the interface restarts or the node |
126 | * reboots). The object cookie only matches one object on that interface | |
4420cfd3 JS |
127 | * during that object's lifetime (i.e. no cookie re-use). |
128 | */ | |
db18b8e9 JS |
129 | typedef struct { |
130 | __u64 wh_interface_cookie; | |
131 | __u64 wh_object_cookie; | |
132 | } WIRE_ATTR lnet_handle_wire_t; | |
133 | ||
134 | typedef enum { | |
135 | LNET_MSG_ACK = 0, | |
136 | LNET_MSG_PUT, | |
137 | LNET_MSG_GET, | |
138 | LNET_MSG_REPLY, | |
139 | LNET_MSG_HELLO, | |
140 | } lnet_msg_type_t; | |
141 | ||
4420cfd3 JS |
142 | /* |
143 | * The variant fields of the portals message header are aligned on an 8 | |
db18b8e9 JS |
144 | * byte boundary in the message header. Note that all types used in these |
145 | * wire structs MUST be fixed size and the smaller types are placed at the | |
4420cfd3 JS |
146 | * end. |
147 | */ | |
db18b8e9 JS |
148 | typedef struct lnet_ack { |
149 | lnet_handle_wire_t dst_wmd; | |
150 | __u64 match_bits; | |
151 | __u32 mlength; | |
152 | } WIRE_ATTR lnet_ack_t; | |
153 | ||
154 | typedef struct lnet_put { | |
155 | lnet_handle_wire_t ack_wmd; | |
156 | __u64 match_bits; | |
157 | __u64 hdr_data; | |
158 | __u32 ptl_index; | |
159 | __u32 offset; | |
160 | } WIRE_ATTR lnet_put_t; | |
161 | ||
162 | typedef struct lnet_get { | |
163 | lnet_handle_wire_t return_wmd; | |
164 | __u64 match_bits; | |
165 | __u32 ptl_index; | |
166 | __u32 src_offset; | |
167 | __u32 sink_length; | |
168 | } WIRE_ATTR lnet_get_t; | |
169 | ||
170 | typedef struct lnet_reply { | |
171 | lnet_handle_wire_t dst_wmd; | |
172 | } WIRE_ATTR lnet_reply_t; | |
173 | ||
174 | typedef struct lnet_hello { | |
175 | __u64 incarnation; | |
176 | __u32 type; | |
177 | } WIRE_ATTR lnet_hello_t; | |
178 | ||
179 | typedef struct { | |
180 | lnet_nid_t dest_nid; | |
181 | lnet_nid_t src_nid; | |
182 | lnet_pid_t dest_pid; | |
183 | lnet_pid_t src_pid; | |
184 | __u32 type; /* lnet_msg_type_t */ | |
185 | __u32 payload_length; /* payload data to follow */ | |
186 | /*<------__u64 aligned------->*/ | |
187 | union { | |
188 | lnet_ack_t ack; | |
189 | lnet_put_t put; | |
190 | lnet_get_t get; | |
191 | lnet_reply_t reply; | |
192 | lnet_hello_t hello; | |
193 | } msg; | |
194 | } WIRE_ATTR lnet_hdr_t; | |
195 | ||
4420cfd3 JS |
196 | /* |
197 | * A HELLO message contains a magic number and protocol version | |
db18b8e9 JS |
198 | * code in the header's dest_nid, the peer's NID in the src_nid, and |
199 | * LNET_MSG_HELLO in the type field. All other common fields are zero | |
200 | * (including payload_size; i.e. no payload). | |
201 | * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is | |
202 | * running the same protocol and to find out its NID. These LNDs should | |
203 | * exchange HELLO messages when a connection is first established. Individual | |
204 | * LNDs can put whatever else they fancy in lnet_hdr_t::msg. | |
205 | */ | |
206 | typedef struct { | |
207 | __u32 magic; /* LNET_PROTO_TCP_MAGIC */ | |
208 | __u16 version_major; /* increment on incompatible change */ | |
209 | __u16 version_minor; /* increment on compatible change */ | |
210 | } WIRE_ATTR lnet_magicversion_t; | |
211 | ||
212 | /* PROTO MAGIC for LNDs */ | |
213 | #define LNET_PROTO_IB_MAGIC 0x0be91b91 | |
214 | #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */ | |
215 | #define LNET_PROTO_TCP_MAGIC 0xeebc0ded | |
216 | #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100 | |
217 | #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */ | |
218 | ||
219 | /* Placeholder for a future "unified" protocol across all LNDs */ | |
4420cfd3 JS |
220 | /* |
221 | * Current LNDs that receive a request with this magic will respond with a | |
222 | * "stub" reply using their current protocol | |
223 | */ | |
db18b8e9 JS |
224 | #define LNET_PROTO_MAGIC 0x45726963 /* ! */ |
225 | ||
226 | #define LNET_PROTO_TCP_VERSION_MAJOR 1 | |
227 | #define LNET_PROTO_TCP_VERSION_MINOR 0 | |
228 | ||
229 | /* Acceptor connection request */ | |
230 | typedef struct { | |
231 | __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */ | |
232 | __u32 acr_version; /* protocol version */ | |
233 | __u64 acr_nid; /* target NID */ | |
234 | } WIRE_ATTR lnet_acceptor_connreq_t; | |
235 | ||
236 | #define LNET_PROTO_ACCEPTOR_VERSION 1 | |
237 | ||
238 | typedef struct { | |
239 | lnet_nid_t ns_nid; | |
240 | __u32 ns_status; | |
241 | __u32 ns_unused; | |
242 | } WIRE_ATTR lnet_ni_status_t; | |
243 | ||
244 | typedef struct { | |
245 | __u32 pi_magic; | |
246 | __u32 pi_features; | |
247 | lnet_pid_t pi_pid; | |
248 | __u32 pi_nnis; | |
249 | lnet_ni_status_t pi_ni[0]; | |
250 | } WIRE_ATTR lnet_ping_info_t; | |
251 | ||
252 | typedef struct lnet_counters { | |
253 | __u32 msgs_alloc; | |
254 | __u32 msgs_max; | |
255 | __u32 errors; | |
256 | __u32 send_count; | |
257 | __u32 recv_count; | |
258 | __u32 route_count; | |
259 | __u32 drop_count; | |
260 | __u64 send_length; | |
261 | __u64 recv_length; | |
262 | __u64 route_length; | |
263 | __u64 drop_length; | |
264 | } WIRE_ATTR lnet_counters_t; | |
265 | ||
266 | #define LNET_NI_STATUS_UP 0x15aac0de | |
267 | #define LNET_NI_STATUS_DOWN 0xdeadface | |
268 | #define LNET_NI_STATUS_INVALID 0x00000000 | |
269 | ||
270 | #define LNET_MAX_INTERFACES 16 | |
271 | ||
4420cfd3 | 272 | /** |
d7e09d03 PT |
273 | * Objects maintained by the LNet are accessed through handles. Handle types |
274 | * have names of the form lnet_handle_xx_t, where xx is one of the two letter | |
275 | * object type codes ('eq' for event queue, 'md' for memory descriptor, and | |
276 | * 'me' for match entry). | |
277 | * Each type of object is given a unique handle type to enhance type checking. | |
278 | * The type lnet_handle_any_t can be used when a generic handle is needed. | |
279 | * Every handle value can be converted into a value of type lnet_handle_any_t | |
280 | * without loss of information. | |
281 | */ | |
282 | typedef struct { | |
283 | __u64 cookie; | |
284 | } lnet_handle_any_t; | |
285 | ||
286 | typedef lnet_handle_any_t lnet_handle_eq_t; | |
287 | typedef lnet_handle_any_t lnet_handle_md_t; | |
288 | typedef lnet_handle_any_t lnet_handle_me_t; | |
289 | ||
290 | #define LNET_WIRE_HANDLE_COOKIE_NONE (-1) | |
291 | ||
292 | /** | |
293 | * Invalidate handle \a h. | |
294 | */ | |
295 | static inline void LNetInvalidateHandle(lnet_handle_any_t *h) | |
296 | { | |
297 | h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE; | |
298 | } | |
299 | ||
300 | /** | |
301 | * Compare handles \a h1 and \a h2. | |
302 | * | |
303 | * \return 1 if handles are equal, 0 if otherwise. | |
304 | */ | |
568414f1 | 305 | static inline int LNetHandleIsEqual(lnet_handle_any_t h1, lnet_handle_any_t h2) |
d7e09d03 | 306 | { |
e4422943 | 307 | return h1.cookie == h2.cookie; |
d7e09d03 PT |
308 | } |
309 | ||
310 | /** | |
311 | * Check whether handle \a h is invalid. | |
312 | * | |
313 | * \return 1 if handle is invalid, 0 if valid. | |
314 | */ | |
315 | static inline int LNetHandleIsInvalid(lnet_handle_any_t h) | |
316 | { | |
625a3121 | 317 | return h.cookie == LNET_WIRE_HANDLE_COOKIE_NONE; |
d7e09d03 PT |
318 | } |
319 | ||
320 | /** | |
321 | * Global process ID. | |
322 | */ | |
323 | typedef struct { | |
324 | /** node id */ | |
325 | lnet_nid_t nid; | |
326 | /** process id */ | |
327 | lnet_pid_t pid; | |
328 | } lnet_process_id_t; | |
329 | /** @} lnet_addr */ | |
330 | ||
331 | /** \addtogroup lnet_me | |
4420cfd3 JS |
332 | * @{ |
333 | */ | |
d7e09d03 PT |
334 | |
335 | /** | |
336 | * Specifies whether the match entry or memory descriptor should be unlinked | |
337 | * automatically (LNET_UNLINK) or not (LNET_RETAIN). | |
338 | */ | |
339 | typedef enum { | |
340 | LNET_RETAIN = 0, | |
341 | LNET_UNLINK | |
342 | } lnet_unlink_t; | |
343 | ||
344 | /** | |
345 | * Values of the type lnet_ins_pos_t are used to control where a new match | |
346 | * entry is inserted. The value LNET_INS_BEFORE is used to insert the new | |
347 | * entry before the current entry or before the head of the list. The value | |
348 | * LNET_INS_AFTER is used to insert the new entry after the current entry | |
349 | * or after the last item in the list. | |
350 | */ | |
351 | typedef enum { | |
352 | /** insert ME before current position or head of the list */ | |
353 | LNET_INS_BEFORE, | |
354 | /** insert ME after current position or tail of the list */ | |
355 | LNET_INS_AFTER, | |
356 | /** attach ME at tail of local CPU partition ME list */ | |
357 | LNET_INS_LOCAL | |
358 | } lnet_ins_pos_t; | |
359 | ||
360 | /** @} lnet_me */ | |
361 | ||
362 | /** \addtogroup lnet_md | |
4420cfd3 JS |
363 | * @{ |
364 | */ | |
d7e09d03 PT |
365 | |
366 | /** | |
367 | * Defines the visible parts of a memory descriptor. Values of this type | |
368 | * are used to initialize memory descriptors. | |
369 | */ | |
370 | typedef struct { | |
371 | /** | |
372 | * Specify the memory region associated with the memory descriptor. | |
373 | * If the options field has: | |
374 | * - LNET_MD_KIOV bit set: The start field points to the starting | |
375 | * address of an array of lnet_kiov_t and the length field specifies | |
376 | * the number of entries in the array. The length can't be bigger | |
377 | * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based | |
d766b4b5 | 378 | * fragments that are not necessarily mapped in virtual memory. |
d7e09d03 PT |
379 | * - LNET_MD_IOVEC bit set: The start field points to the starting |
380 | * address of an array of struct iovec and the length field specifies | |
381 | * the number of entries in the array. The length can't be bigger | |
382 | * than LNET_MAX_IOV. The struct iovec is used to describe fragments | |
383 | * that have virtual addresses. | |
384 | * - Otherwise: The memory region is contiguous. The start field | |
385 | * specifies the starting address for the memory region and the | |
386 | * length field specifies its length. | |
387 | * | |
388 | * When the memory region is fragmented, all fragments but the first | |
389 | * one must start on page boundary, and all but the last must end on | |
390 | * page boundary. | |
391 | */ | |
188acc61 JS |
392 | void *start; |
393 | unsigned int length; | |
d7e09d03 PT |
394 | /** |
395 | * Specifies the maximum number of operations that can be performed | |
396 | * on the memory descriptor. An operation is any action that could | |
397 | * possibly generate an event. In the usual case, the threshold value | |
398 | * is decremented for each operation on the MD. When the threshold | |
399 | * drops to zero, the MD becomes inactive and does not respond to | |
400 | * operations. A threshold value of LNET_MD_THRESH_INF indicates that | |
401 | * there is no bound on the number of operations that may be applied | |
402 | * to a MD. | |
403 | */ | |
188acc61 | 404 | int threshold; |
d7e09d03 PT |
405 | /** |
406 | * Specifies the largest incoming request that the memory descriptor | |
407 | * should respond to. When the unused portion of a MD (length - | |
408 | * local offset) falls below this value, the MD becomes inactive and | |
409 | * does not respond to further operations. This value is only used | |
410 | * if the LNET_MD_MAX_SIZE option is set. | |
411 | */ | |
188acc61 | 412 | int max_size; |
d7e09d03 PT |
413 | /** |
414 | * Specifies the behavior of the memory descriptor. A bitwise OR | |
415 | * of the following values can be used: | |
416 | * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD. | |
417 | * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD. | |
418 | * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory | |
419 | * region is provided by the incoming request. By default, the | |
420 | * offset is maintained locally. When maintained locally, the | |
421 | * offset is incremented by the length of the request so that | |
422 | * the next operation (PUT or GET) will access the next part of | |
423 | * the memory region. Note that only one offset variable exists | |
424 | * per memory descriptor. If both PUT and GET operations are | |
425 | * performed on a memory descriptor, the offset is updated each time. | |
426 | * - LNET_MD_TRUNCATE: The length provided in the incoming request can | |
427 | * be reduced to match the memory available in the region (determined | |
428 | * by subtracting the offset from the length of the memory region). | |
429 | * By default, if the length in the incoming operation is greater | |
430 | * than the amount of memory available, the operation is rejected. | |
431 | * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for | |
432 | * incoming PUT operations, even if requested. By default, | |
433 | * acknowledgments are sent for PUT operations that request an | |
434 | * acknowledgment. Acknowledgments are never sent for GET operations. | |
435 | * The data sent in the REPLY serves as an implicit acknowledgment. | |
436 | * - LNET_MD_KIOV: The start and length fields specify an array of | |
437 | * lnet_kiov_t. | |
438 | * - LNET_MD_IOVEC: The start and length fields specify an array of | |
439 | * struct iovec. | |
440 | * - LNET_MD_MAX_SIZE: The max_size field is valid. | |
441 | * | |
442 | * Note: | |
443 | * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather | |
444 | * capability for memory descriptors. They can't be both set. | |
445 | * - When LNET_MD_MAX_SIZE is set, the total length of the memory | |
446 | * region (i.e. sum of all fragment lengths) must not be less than | |
447 | * \a max_size. | |
448 | */ | |
188acc61 | 449 | unsigned int options; |
d7e09d03 PT |
450 | /** |
451 | * A user-specified value that is associated with the memory | |
452 | * descriptor. The value does not need to be a pointer, but must fit | |
453 | * in the space used by a pointer. This value is recorded in events | |
454 | * associated with operations on this MD. | |
455 | */ | |
188acc61 | 456 | void *user_ptr; |
d7e09d03 PT |
457 | /** |
458 | * A handle for the event queue used to log the operations performed on | |
459 | * the memory region. If this argument is a NULL handle (i.e. nullified | |
460 | * by LNetInvalidateHandle()), operations performed on this memory | |
461 | * descriptor are not logged. | |
462 | */ | |
463 | lnet_handle_eq_t eq_handle; | |
464 | } lnet_md_t; | |
465 | ||
4420cfd3 JS |
466 | /* |
467 | * Max Transfer Unit (minimum supported everywhere). | |
d7e09d03 | 468 | * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks) |
4420cfd3 JS |
469 | * these limits are system wide and not interface-local. |
470 | */ | |
d7e09d03 PT |
471 | #define LNET_MTU_BITS 20 |
472 | #define LNET_MTU (1 << LNET_MTU_BITS) | |
473 | ||
474 | /** limit on the number of fragments in discontiguous MDs */ | |
188acc61 | 475 | #define LNET_MAX_IOV 256 |
d7e09d03 | 476 | |
d7e09d03 PT |
477 | /** |
478 | * Options for the MD structure. See lnet_md_t::options. | |
479 | */ | |
188acc61 | 480 | #define LNET_MD_OP_PUT (1 << 0) |
d7e09d03 | 481 | /** See lnet_md_t::options. */ |
188acc61 | 482 | #define LNET_MD_OP_GET (1 << 1) |
d7e09d03 PT |
483 | /** See lnet_md_t::options. */ |
484 | #define LNET_MD_MANAGE_REMOTE (1 << 2) | |
188acc61 | 485 | /* unused (1 << 3) */ |
d7e09d03 | 486 | /** See lnet_md_t::options. */ |
188acc61 | 487 | #define LNET_MD_TRUNCATE (1 << 4) |
d7e09d03 | 488 | /** See lnet_md_t::options. */ |
188acc61 | 489 | #define LNET_MD_ACK_DISABLE (1 << 5) |
d7e09d03 PT |
490 | /** See lnet_md_t::options. */ |
491 | #define LNET_MD_IOVEC (1 << 6) | |
492 | /** See lnet_md_t::options. */ | |
188acc61 | 493 | #define LNET_MD_MAX_SIZE (1 << 7) |
d7e09d03 | 494 | /** See lnet_md_t::options. */ |
188acc61 | 495 | #define LNET_MD_KIOV (1 << 8) |
d7e09d03 PT |
496 | |
497 | /* For compatibility with Cray Portals */ | |
188acc61 | 498 | #define LNET_MD_PHYS 0 |
d7e09d03 PT |
499 | |
500 | /** Infinite threshold on MD operations. See lnet_md_t::threshold */ | |
188acc61 | 501 | #define LNET_MD_THRESH_INF (-1) |
d7e09d03 PT |
502 | |
503 | /* NB lustre portals uses struct iovec internally! */ | |
504 | typedef struct iovec lnet_md_iovec_t; | |
505 | ||
506 | /** | |
507 | * A page-based fragment of a MD. | |
508 | */ | |
509 | typedef struct { | |
510 | /** Pointer to the page where the fragment resides */ | |
188acc61 | 511 | struct page *kiov_page; |
d7e09d03 | 512 | /** Length in bytes of the fragment */ |
188acc61 | 513 | unsigned int kiov_len; |
d7e09d03 PT |
514 | /** |
515 | * Starting offset of the fragment within the page. Note that the | |
516 | * end of the fragment must not pass the end of the page; i.e., | |
ea1754a0 | 517 | * kiov_len + kiov_offset <= PAGE_SIZE. |
d7e09d03 | 518 | */ |
188acc61 | 519 | unsigned int kiov_offset; |
d7e09d03 PT |
520 | } lnet_kiov_t; |
521 | /** @} lnet_md */ | |
522 | ||
523 | /** \addtogroup lnet_eq | |
4420cfd3 JS |
524 | * @{ |
525 | */ | |
d7e09d03 PT |
526 | |
527 | /** | |
528 | * Six types of events can be logged in an event queue. | |
529 | */ | |
530 | typedef enum { | |
531 | /** An incoming GET operation has completed on the MD. */ | |
532 | LNET_EVENT_GET = 1, | |
533 | /** | |
534 | * An incoming PUT operation has completed on the MD. The | |
535 | * underlying layers will not alter the memory (on behalf of this | |
536 | * operation) once this event has been logged. | |
537 | */ | |
538 | LNET_EVENT_PUT, | |
539 | /** | |
540 | * A REPLY operation has completed. This event is logged after the | |
541 | * data (if any) from the REPLY has been written into the MD. | |
542 | */ | |
543 | LNET_EVENT_REPLY, | |
544 | /** An acknowledgment has been received. */ | |
545 | LNET_EVENT_ACK, | |
546 | /** | |
547 | * An outgoing send (PUT or GET) operation has completed. This event | |
548 | * is logged after the entire buffer has been sent and it is safe for | |
549 | * the caller to reuse the buffer. | |
550 | * | |
551 | * Note: | |
552 | * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can | |
553 | * happen even when the message has not yet been put out on wire. | |
554 | * - It's unsafe to assume that in an outgoing GET operation | |
555 | * the LNET_EVENT_SEND event would happen before the | |
556 | * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and | |
557 | * LNET_EVENT_ACK events in an outgoing PUT operation. | |
558 | */ | |
559 | LNET_EVENT_SEND, | |
560 | /** | |
561 | * A MD has been unlinked. Note that LNetMDUnlink() does not | |
562 | * necessarily trigger an LNET_EVENT_UNLINK event. | |
563 | * \see LNetMDUnlink | |
564 | */ | |
565 | LNET_EVENT_UNLINK, | |
566 | } lnet_event_kind_t; | |
567 | ||
188acc61 | 568 | #define LNET_SEQ_BASETYPE long |
d7e09d03 | 569 | typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t; |
e4422943 | 570 | #define LNET_SEQ_GT(a, b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0) |
d7e09d03 | 571 | |
d7e09d03 PT |
572 | /** |
573 | * Information about an event on a MD. | |
574 | */ | |
575 | typedef struct { | |
576 | /** The identifier (nid, pid) of the target. */ | |
188acc61 | 577 | lnet_process_id_t target; |
d7e09d03 | 578 | /** The identifier (nid, pid) of the initiator. */ |
188acc61 | 579 | lnet_process_id_t initiator; |
d7e09d03 PT |
580 | /** |
581 | * The NID of the immediate sender. If the request has been forwarded | |
582 | * by routers, this is the NID of the last hop; otherwise it's the | |
583 | * same as the initiator. | |
584 | */ | |
188acc61 | 585 | lnet_nid_t sender; |
d7e09d03 | 586 | /** Indicates the type of the event. */ |
188acc61 | 587 | lnet_event_kind_t type; |
d7e09d03 | 588 | /** The portal table index specified in the request */ |
188acc61 | 589 | unsigned int pt_index; |
d7e09d03 | 590 | /** A copy of the match bits specified in the request. */ |
188acc61 | 591 | __u64 match_bits; |
d7e09d03 | 592 | /** The length (in bytes) specified in the request. */ |
188acc61 | 593 | unsigned int rlength; |
d7e09d03 PT |
594 | /** |
595 | * The length (in bytes) of the data that was manipulated by the | |
596 | * operation. For truncated operations, the manipulated length will be | |
597 | * the number of bytes specified by the MD (possibly with an offset, | |
598 | * see lnet_md_t). For all other operations, the manipulated length | |
599 | * will be the length of the requested operation, i.e. rlength. | |
600 | */ | |
188acc61 | 601 | unsigned int mlength; |
d7e09d03 PT |
602 | /** |
603 | * The handle to the MD associated with the event. The handle may be | |
604 | * invalid if the MD has been unlinked. | |
605 | */ | |
188acc61 | 606 | lnet_handle_md_t md_handle; |
d7e09d03 PT |
607 | /** |
608 | * A snapshot of the state of the MD immediately after the event has | |
609 | * been processed. In particular, the threshold field in md will | |
610 | * reflect the value of the threshold after the operation occurred. | |
611 | */ | |
188acc61 | 612 | lnet_md_t md; |
d7e09d03 PT |
613 | /** |
614 | * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT. | |
615 | * \see LNetPut | |
616 | */ | |
188acc61 | 617 | __u64 hdr_data; |
d7e09d03 PT |
618 | /** |
619 | * Indicates the completion status of the operation. It's 0 for | |
620 | * successful operations, otherwise it's an error code. | |
621 | */ | |
188acc61 | 622 | int status; |
d7e09d03 PT |
623 | /** |
624 | * Indicates whether the MD has been unlinked. Note that: | |
625 | * - An event with unlinked set is the last event on the MD. | |
626 | * - This field is also set for an explicit LNET_EVENT_UNLINK event. | |
627 | * \see LNetMDUnlink | |
628 | */ | |
188acc61 | 629 | int unlinked; |
d7e09d03 PT |
630 | /** |
631 | * The displacement (in bytes) into the memory region that the | |
632 | * operation used. The offset can be determined by the operation for | |
633 | * a remote managed MD or by the local MD. | |
634 | * \see lnet_md_t::options | |
635 | */ | |
188acc61 | 636 | unsigned int offset; |
d7e09d03 PT |
637 | /** |
638 | * The sequence number for this event. Sequence numbers are unique | |
639 | * to each event. | |
640 | */ | |
188acc61 | 641 | volatile lnet_seq_t sequence; |
d7e09d03 | 642 | } lnet_event_t; |
d7e09d03 PT |
643 | |
644 | /** | |
645 | * Event queue handler function type. | |
646 | * | |
647 | * The EQ handler runs for each event that is deposited into the EQ. The | |
648 | * handler is supplied with a pointer to the event that triggered the | |
649 | * handler invocation. | |
650 | * | |
651 | * The handler must not block, must be reentrant, and must not call any LNet | |
652 | * API functions. It should return as quickly as possible. | |
653 | */ | |
654 | typedef void (*lnet_eq_handler_t)(lnet_event_t *event); | |
655 | #define LNET_EQ_HANDLER_NONE NULL | |
656 | /** @} lnet_eq */ | |
657 | ||
658 | /** \addtogroup lnet_data | |
4420cfd3 JS |
659 | * @{ |
660 | */ | |
d7e09d03 PT |
661 | |
662 | /** | |
663 | * Specify whether an acknowledgment should be sent by target when the PUT | |
664 | * operation completes (i.e., when the data has been written to a MD of the | |
665 | * target process). | |
666 | * | |
667 | * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which | |
668 | * acknowledgments can be disabled for a MD. | |
669 | */ | |
670 | typedef enum { | |
671 | /** Request an acknowledgment */ | |
672 | LNET_ACK_REQ, | |
673 | /** Request that no acknowledgment should be generated. */ | |
674 | LNET_NOACK_REQ | |
675 | } lnet_ack_req_t; | |
676 | /** @} lnet_data */ | |
677 | ||
678 | /** @} lnet */ | |
679 | #endif |