Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 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 | |
| 40 | /** \addtogroup lnet |
| 41 | * @{ */ |
| 42 | |
| 43 | #include <linux/libcfs/libcfs.h> |
| 44 | |
| 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 */ |
| 77 | |
| 78 | #define LNET_TIME_FOREVER (-1) |
| 79 | |
| 80 | /** |
| 81 | * Objects maintained by the LNet are accessed through handles. Handle types |
| 82 | * have names of the form lnet_handle_xx_t, where xx is one of the two letter |
| 83 | * object type codes ('eq' for event queue, 'md' for memory descriptor, and |
| 84 | * 'me' for match entry). |
| 85 | * Each type of object is given a unique handle type to enhance type checking. |
| 86 | * The type lnet_handle_any_t can be used when a generic handle is needed. |
| 87 | * Every handle value can be converted into a value of type lnet_handle_any_t |
| 88 | * without loss of information. |
| 89 | */ |
| 90 | typedef struct { |
| 91 | __u64 cookie; |
| 92 | } lnet_handle_any_t; |
| 93 | |
| 94 | typedef lnet_handle_any_t lnet_handle_eq_t; |
| 95 | typedef lnet_handle_any_t lnet_handle_md_t; |
| 96 | typedef lnet_handle_any_t lnet_handle_me_t; |
| 97 | |
| 98 | #define LNET_WIRE_HANDLE_COOKIE_NONE (-1) |
| 99 | |
| 100 | /** |
| 101 | * Invalidate handle \a h. |
| 102 | */ |
| 103 | static inline void LNetInvalidateHandle(lnet_handle_any_t *h) |
| 104 | { |
| 105 | h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE; |
| 106 | } |
| 107 | |
| 108 | /** |
| 109 | * Compare handles \a h1 and \a h2. |
| 110 | * |
| 111 | * \return 1 if handles are equal, 0 if otherwise. |
| 112 | */ |
| 113 | static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2) |
| 114 | { |
| 115 | return (h1.cookie == h2.cookie); |
| 116 | } |
| 117 | |
| 118 | /** |
| 119 | * Check whether handle \a h is invalid. |
| 120 | * |
| 121 | * \return 1 if handle is invalid, 0 if valid. |
| 122 | */ |
| 123 | static inline int LNetHandleIsInvalid(lnet_handle_any_t h) |
| 124 | { |
| 125 | return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie); |
| 126 | } |
| 127 | |
| 128 | /** |
| 129 | * Global process ID. |
| 130 | */ |
| 131 | typedef struct { |
| 132 | /** node id */ |
| 133 | lnet_nid_t nid; |
| 134 | /** process id */ |
| 135 | lnet_pid_t pid; |
| 136 | } lnet_process_id_t; |
| 137 | /** @} lnet_addr */ |
| 138 | |
| 139 | /** \addtogroup lnet_me |
| 140 | * @{ */ |
| 141 | |
| 142 | /** |
| 143 | * Specifies whether the match entry or memory descriptor should be unlinked |
| 144 | * automatically (LNET_UNLINK) or not (LNET_RETAIN). |
| 145 | */ |
| 146 | typedef enum { |
| 147 | LNET_RETAIN = 0, |
| 148 | LNET_UNLINK |
| 149 | } lnet_unlink_t; |
| 150 | |
| 151 | /** |
| 152 | * Values of the type lnet_ins_pos_t are used to control where a new match |
| 153 | * entry is inserted. The value LNET_INS_BEFORE is used to insert the new |
| 154 | * entry before the current entry or before the head of the list. The value |
| 155 | * LNET_INS_AFTER is used to insert the new entry after the current entry |
| 156 | * or after the last item in the list. |
| 157 | */ |
| 158 | typedef enum { |
| 159 | /** insert ME before current position or head of the list */ |
| 160 | LNET_INS_BEFORE, |
| 161 | /** insert ME after current position or tail of the list */ |
| 162 | LNET_INS_AFTER, |
| 163 | /** attach ME at tail of local CPU partition ME list */ |
| 164 | LNET_INS_LOCAL |
| 165 | } lnet_ins_pos_t; |
| 166 | |
| 167 | /** @} lnet_me */ |
| 168 | |
| 169 | /** \addtogroup lnet_md |
| 170 | * @{ */ |
| 171 | |
| 172 | /** |
| 173 | * Defines the visible parts of a memory descriptor. Values of this type |
| 174 | * are used to initialize memory descriptors. |
| 175 | */ |
| 176 | typedef struct { |
| 177 | /** |
| 178 | * Specify the memory region associated with the memory descriptor. |
| 179 | * If the options field has: |
| 180 | * - LNET_MD_KIOV bit set: The start field points to the starting |
| 181 | * address of an array of lnet_kiov_t and the length field specifies |
| 182 | * the number of entries in the array. The length can't be bigger |
| 183 | * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based |
Masanari Iida | d766b4b | 2014-03-21 23:42:52 +0900 | [diff] [blame^] | 184 | * fragments that are not necessarily mapped in virtual memory. |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 185 | * - LNET_MD_IOVEC bit set: The start field points to the starting |
| 186 | * address of an array of struct iovec and the length field specifies |
| 187 | * the number of entries in the array. The length can't be bigger |
| 188 | * than LNET_MAX_IOV. The struct iovec is used to describe fragments |
| 189 | * that have virtual addresses. |
| 190 | * - Otherwise: The memory region is contiguous. The start field |
| 191 | * specifies the starting address for the memory region and the |
| 192 | * length field specifies its length. |
| 193 | * |
| 194 | * When the memory region is fragmented, all fragments but the first |
| 195 | * one must start on page boundary, and all but the last must end on |
| 196 | * page boundary. |
| 197 | */ |
| 198 | void *start; |
| 199 | unsigned int length; |
| 200 | /** |
| 201 | * Specifies the maximum number of operations that can be performed |
| 202 | * on the memory descriptor. An operation is any action that could |
| 203 | * possibly generate an event. In the usual case, the threshold value |
| 204 | * is decremented for each operation on the MD. When the threshold |
| 205 | * drops to zero, the MD becomes inactive and does not respond to |
| 206 | * operations. A threshold value of LNET_MD_THRESH_INF indicates that |
| 207 | * there is no bound on the number of operations that may be applied |
| 208 | * to a MD. |
| 209 | */ |
| 210 | int threshold; |
| 211 | /** |
| 212 | * Specifies the largest incoming request that the memory descriptor |
| 213 | * should respond to. When the unused portion of a MD (length - |
| 214 | * local offset) falls below this value, the MD becomes inactive and |
| 215 | * does not respond to further operations. This value is only used |
| 216 | * if the LNET_MD_MAX_SIZE option is set. |
| 217 | */ |
| 218 | int max_size; |
| 219 | /** |
| 220 | * Specifies the behavior of the memory descriptor. A bitwise OR |
| 221 | * of the following values can be used: |
| 222 | * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD. |
| 223 | * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD. |
| 224 | * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory |
| 225 | * region is provided by the incoming request. By default, the |
| 226 | * offset is maintained locally. When maintained locally, the |
| 227 | * offset is incremented by the length of the request so that |
| 228 | * the next operation (PUT or GET) will access the next part of |
| 229 | * the memory region. Note that only one offset variable exists |
| 230 | * per memory descriptor. If both PUT and GET operations are |
| 231 | * performed on a memory descriptor, the offset is updated each time. |
| 232 | * - LNET_MD_TRUNCATE: The length provided in the incoming request can |
| 233 | * be reduced to match the memory available in the region (determined |
| 234 | * by subtracting the offset from the length of the memory region). |
| 235 | * By default, if the length in the incoming operation is greater |
| 236 | * than the amount of memory available, the operation is rejected. |
| 237 | * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for |
| 238 | * incoming PUT operations, even if requested. By default, |
| 239 | * acknowledgments are sent for PUT operations that request an |
| 240 | * acknowledgment. Acknowledgments are never sent for GET operations. |
| 241 | * The data sent in the REPLY serves as an implicit acknowledgment. |
| 242 | * - LNET_MD_KIOV: The start and length fields specify an array of |
| 243 | * lnet_kiov_t. |
| 244 | * - LNET_MD_IOVEC: The start and length fields specify an array of |
| 245 | * struct iovec. |
| 246 | * - LNET_MD_MAX_SIZE: The max_size field is valid. |
| 247 | * |
| 248 | * Note: |
| 249 | * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather |
| 250 | * capability for memory descriptors. They can't be both set. |
| 251 | * - When LNET_MD_MAX_SIZE is set, the total length of the memory |
| 252 | * region (i.e. sum of all fragment lengths) must not be less than |
| 253 | * \a max_size. |
| 254 | */ |
| 255 | unsigned int options; |
| 256 | /** |
| 257 | * A user-specified value that is associated with the memory |
| 258 | * descriptor. The value does not need to be a pointer, but must fit |
| 259 | * in the space used by a pointer. This value is recorded in events |
| 260 | * associated with operations on this MD. |
| 261 | */ |
| 262 | void *user_ptr; |
| 263 | /** |
| 264 | * A handle for the event queue used to log the operations performed on |
| 265 | * the memory region. If this argument is a NULL handle (i.e. nullified |
| 266 | * by LNetInvalidateHandle()), operations performed on this memory |
| 267 | * descriptor are not logged. |
| 268 | */ |
| 269 | lnet_handle_eq_t eq_handle; |
| 270 | } lnet_md_t; |
| 271 | |
| 272 | /* Max Transfer Unit (minimum supported everywhere). |
| 273 | * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks) |
| 274 | * these limits are system wide and not interface-local. */ |
| 275 | #define LNET_MTU_BITS 20 |
| 276 | #define LNET_MTU (1 << LNET_MTU_BITS) |
| 277 | |
| 278 | /** limit on the number of fragments in discontiguous MDs */ |
| 279 | #define LNET_MAX_IOV 256 |
| 280 | |
| 281 | /* Max payload size */ |
| 282 | # define LNET_MAX_PAYLOAD CONFIG_LNET_MAX_PAYLOAD |
| 283 | # if (LNET_MAX_PAYLOAD < LNET_MTU) |
| 284 | # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb" |
| 285 | # else |
| 286 | # if (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV)) |
| 287 | /* PAGE_SIZE is a constant: check with cpp! */ |
| 288 | # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb" |
| 289 | # endif |
| 290 | # endif |
| 291 | |
| 292 | /** |
| 293 | * Options for the MD structure. See lnet_md_t::options. |
| 294 | */ |
| 295 | #define LNET_MD_OP_PUT (1 << 0) |
| 296 | /** See lnet_md_t::options. */ |
| 297 | #define LNET_MD_OP_GET (1 << 1) |
| 298 | /** See lnet_md_t::options. */ |
| 299 | #define LNET_MD_MANAGE_REMOTE (1 << 2) |
| 300 | /* unused (1 << 3) */ |
| 301 | /** See lnet_md_t::options. */ |
| 302 | #define LNET_MD_TRUNCATE (1 << 4) |
| 303 | /** See lnet_md_t::options. */ |
| 304 | #define LNET_MD_ACK_DISABLE (1 << 5) |
| 305 | /** See lnet_md_t::options. */ |
| 306 | #define LNET_MD_IOVEC (1 << 6) |
| 307 | /** See lnet_md_t::options. */ |
| 308 | #define LNET_MD_MAX_SIZE (1 << 7) |
| 309 | /** See lnet_md_t::options. */ |
| 310 | #define LNET_MD_KIOV (1 << 8) |
| 311 | |
| 312 | /* For compatibility with Cray Portals */ |
| 313 | #define LNET_MD_PHYS 0 |
| 314 | |
| 315 | /** Infinite threshold on MD operations. See lnet_md_t::threshold */ |
| 316 | #define LNET_MD_THRESH_INF (-1) |
| 317 | |
| 318 | /* NB lustre portals uses struct iovec internally! */ |
| 319 | typedef struct iovec lnet_md_iovec_t; |
| 320 | |
| 321 | /** |
| 322 | * A page-based fragment of a MD. |
| 323 | */ |
| 324 | typedef struct { |
| 325 | /** Pointer to the page where the fragment resides */ |
| 326 | struct page *kiov_page; |
| 327 | /** Length in bytes of the fragment */ |
| 328 | unsigned int kiov_len; |
| 329 | /** |
| 330 | * Starting offset of the fragment within the page. Note that the |
| 331 | * end of the fragment must not pass the end of the page; i.e., |
| 332 | * kiov_len + kiov_offset <= PAGE_CACHE_SIZE. |
| 333 | */ |
| 334 | unsigned int kiov_offset; |
| 335 | } lnet_kiov_t; |
| 336 | /** @} lnet_md */ |
| 337 | |
| 338 | /** \addtogroup lnet_eq |
| 339 | * @{ */ |
| 340 | |
| 341 | /** |
| 342 | * Six types of events can be logged in an event queue. |
| 343 | */ |
| 344 | typedef enum { |
| 345 | /** An incoming GET operation has completed on the MD. */ |
| 346 | LNET_EVENT_GET = 1, |
| 347 | /** |
| 348 | * An incoming PUT operation has completed on the MD. The |
| 349 | * underlying layers will not alter the memory (on behalf of this |
| 350 | * operation) once this event has been logged. |
| 351 | */ |
| 352 | LNET_EVENT_PUT, |
| 353 | /** |
| 354 | * A REPLY operation has completed. This event is logged after the |
| 355 | * data (if any) from the REPLY has been written into the MD. |
| 356 | */ |
| 357 | LNET_EVENT_REPLY, |
| 358 | /** An acknowledgment has been received. */ |
| 359 | LNET_EVENT_ACK, |
| 360 | /** |
| 361 | * An outgoing send (PUT or GET) operation has completed. This event |
| 362 | * is logged after the entire buffer has been sent and it is safe for |
| 363 | * the caller to reuse the buffer. |
| 364 | * |
| 365 | * Note: |
| 366 | * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can |
| 367 | * happen even when the message has not yet been put out on wire. |
| 368 | * - It's unsafe to assume that in an outgoing GET operation |
| 369 | * the LNET_EVENT_SEND event would happen before the |
| 370 | * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and |
| 371 | * LNET_EVENT_ACK events in an outgoing PUT operation. |
| 372 | */ |
| 373 | LNET_EVENT_SEND, |
| 374 | /** |
| 375 | * A MD has been unlinked. Note that LNetMDUnlink() does not |
| 376 | * necessarily trigger an LNET_EVENT_UNLINK event. |
| 377 | * \see LNetMDUnlink |
| 378 | */ |
| 379 | LNET_EVENT_UNLINK, |
| 380 | } lnet_event_kind_t; |
| 381 | |
| 382 | #define LNET_SEQ_BASETYPE long |
| 383 | typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t; |
| 384 | #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0) |
| 385 | |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 386 | /** |
| 387 | * Information about an event on a MD. |
| 388 | */ |
| 389 | typedef struct { |
| 390 | /** The identifier (nid, pid) of the target. */ |
| 391 | lnet_process_id_t target; |
| 392 | /** The identifier (nid, pid) of the initiator. */ |
| 393 | lnet_process_id_t initiator; |
| 394 | /** |
| 395 | * The NID of the immediate sender. If the request has been forwarded |
| 396 | * by routers, this is the NID of the last hop; otherwise it's the |
| 397 | * same as the initiator. |
| 398 | */ |
| 399 | lnet_nid_t sender; |
| 400 | /** Indicates the type of the event. */ |
| 401 | lnet_event_kind_t type; |
| 402 | /** The portal table index specified in the request */ |
| 403 | unsigned int pt_index; |
| 404 | /** A copy of the match bits specified in the request. */ |
| 405 | __u64 match_bits; |
| 406 | /** The length (in bytes) specified in the request. */ |
| 407 | unsigned int rlength; |
| 408 | /** |
| 409 | * The length (in bytes) of the data that was manipulated by the |
| 410 | * operation. For truncated operations, the manipulated length will be |
| 411 | * the number of bytes specified by the MD (possibly with an offset, |
| 412 | * see lnet_md_t). For all other operations, the manipulated length |
| 413 | * will be the length of the requested operation, i.e. rlength. |
| 414 | */ |
| 415 | unsigned int mlength; |
| 416 | /** |
| 417 | * The handle to the MD associated with the event. The handle may be |
| 418 | * invalid if the MD has been unlinked. |
| 419 | */ |
| 420 | lnet_handle_md_t md_handle; |
| 421 | /** |
| 422 | * A snapshot of the state of the MD immediately after the event has |
| 423 | * been processed. In particular, the threshold field in md will |
| 424 | * reflect the value of the threshold after the operation occurred. |
| 425 | */ |
| 426 | lnet_md_t md; |
| 427 | /** |
| 428 | * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT. |
| 429 | * \see LNetPut |
| 430 | */ |
| 431 | __u64 hdr_data; |
| 432 | /** |
| 433 | * Indicates the completion status of the operation. It's 0 for |
| 434 | * successful operations, otherwise it's an error code. |
| 435 | */ |
| 436 | int status; |
| 437 | /** |
| 438 | * Indicates whether the MD has been unlinked. Note that: |
| 439 | * - An event with unlinked set is the last event on the MD. |
| 440 | * - This field is also set for an explicit LNET_EVENT_UNLINK event. |
| 441 | * \see LNetMDUnlink |
| 442 | */ |
| 443 | int unlinked; |
| 444 | /** |
| 445 | * The displacement (in bytes) into the memory region that the |
| 446 | * operation used. The offset can be determined by the operation for |
| 447 | * a remote managed MD or by the local MD. |
| 448 | * \see lnet_md_t::options |
| 449 | */ |
| 450 | unsigned int offset; |
| 451 | /** |
| 452 | * The sequence number for this event. Sequence numbers are unique |
| 453 | * to each event. |
| 454 | */ |
| 455 | volatile lnet_seq_t sequence; |
| 456 | } lnet_event_t; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 457 | |
| 458 | /** |
| 459 | * Event queue handler function type. |
| 460 | * |
| 461 | * The EQ handler runs for each event that is deposited into the EQ. The |
| 462 | * handler is supplied with a pointer to the event that triggered the |
| 463 | * handler invocation. |
| 464 | * |
| 465 | * The handler must not block, must be reentrant, and must not call any LNet |
| 466 | * API functions. It should return as quickly as possible. |
| 467 | */ |
| 468 | typedef void (*lnet_eq_handler_t)(lnet_event_t *event); |
| 469 | #define LNET_EQ_HANDLER_NONE NULL |
| 470 | /** @} lnet_eq */ |
| 471 | |
| 472 | /** \addtogroup lnet_data |
| 473 | * @{ */ |
| 474 | |
| 475 | /** |
| 476 | * Specify whether an acknowledgment should be sent by target when the PUT |
| 477 | * operation completes (i.e., when the data has been written to a MD of the |
| 478 | * target process). |
| 479 | * |
| 480 | * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which |
| 481 | * acknowledgments can be disabled for a MD. |
| 482 | */ |
| 483 | typedef enum { |
| 484 | /** Request an acknowledgment */ |
| 485 | LNET_ACK_REQ, |
| 486 | /** Request that no acknowledgment should be generated. */ |
| 487 | LNET_NOACK_REQ |
| 488 | } lnet_ack_req_t; |
| 489 | /** @} lnet_data */ |
| 490 | |
| 491 | /** @} lnet */ |
| 492 | #endif |