4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
35 #ifndef __LNET_TYPES_H__
36 #define __LNET_TYPES_H__
41 #include <libcfs/libcfs.h>
43 /** \addtogroup lnet_addr
46 /** Portal reserved for LNet's own use.
47 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
49 #define LNET_RESERVED_PORTAL 0
52 * Address of an end-point in an LNet network.
54 * A node can have multiple end-points and hence multiple addresses.
55 * An LNet network can be a simple network (e.g. tcp0) or a network of
56 * LNet networks connected by LNet routers. Therefore an end-point address
57 * has two parts: network ID, and address within a network.
59 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
61 typedef __u64 lnet_nid_t;
63 * ID of a process in a node. Shortened as PID to distinguish from
64 * lnet_process_id_t, the global process ID.
66 typedef __u32 lnet_pid_t;
68 /** wildcard NID that matches any end-point address */
69 #define LNET_NID_ANY ((lnet_nid_t) -1)
70 /** wildcard PID that matches any lnet_pid_t */
71 #define LNET_PID_ANY ((lnet_pid_t) -1)
73 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
74 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
76 #define LNET_TIME_FOREVER (-1)
79 * Objects maintained by the LNet are accessed through handles. Handle types
80 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
81 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
82 * 'me' for match entry).
83 * Each type of object is given a unique handle type to enhance type checking.
84 * The type lnet_handle_any_t can be used when a generic handle is needed.
85 * Every handle value can be converted into a value of type lnet_handle_any_t
86 * without loss of information.
92 typedef lnet_handle_any_t lnet_handle_eq_t;
93 typedef lnet_handle_any_t lnet_handle_md_t;
94 typedef lnet_handle_any_t lnet_handle_me_t;
96 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
99 * Invalidate handle \a h.
101 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
103 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
107 * Compare handles \a h1 and \a h2.
109 * \return 1 if handles are equal, 0 if otherwise.
111 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
113 return (h1.cookie == h2.cookie);
117 * Check whether handle \a h is invalid.
119 * \return 1 if handle is invalid, 0 if valid.
121 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
123 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
137 /** \addtogroup lnet_me
141 * Specifies whether the match entry or memory descriptor should be unlinked
142 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
150 * Values of the type lnet_ins_pos_t are used to control where a new match
151 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
152 * entry before the current entry or before the head of the list. The value
153 * LNET_INS_AFTER is used to insert the new entry after the current entry
154 * or after the last item in the list.
157 /** insert ME before current position or head of the list */
159 /** insert ME after current position or tail of the list */
161 /** attach ME at tail of local CPU partition ME list */
167 /** \addtogroup lnet_md
171 * Defines the visible parts of a memory descriptor. Values of this type
172 * are used to initialize memory descriptors.
176 * Specify the memory region associated with the memory descriptor.
177 * If the options field has:
178 * - LNET_MD_KIOV bit set: The start field points to the starting
179 * address of an array of lnet_kiov_t and the length field specifies
180 * the number of entries in the array. The length can't be bigger
181 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
182 * fragments that are not necessarily mapped in virtal memory.
183 * - LNET_MD_IOVEC bit set: The start field points to the starting
184 * address of an array of struct iovec and the length field specifies
185 * the number of entries in the array. The length can't be bigger
186 * than LNET_MAX_IOV. The struct iovec is used to describe fragments
187 * that have virtual addresses.
188 * - Otherwise: The memory region is contiguous. The start field
189 * specifies the starting address for the memory region and the
190 * length field specifies its length.
192 * When the memory region is fragmented, all fragments but the first
193 * one must start on page boundary, and all but the last must end on
199 * Specifies the maximum number of operations that can be performed
200 * on the memory descriptor. An operation is any action that could
201 * possibly generate an event. In the usual case, the threshold value
202 * is decremented for each operation on the MD. When the threshold
203 * drops to zero, the MD becomes inactive and does not respond to
204 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
205 * there is no bound on the number of operations that may be applied
210 * Specifies the largest incoming request that the memory descriptor
211 * should respond to. When the unused portion of a MD (length -
212 * local offset) falls below this value, the MD becomes inactive and
213 * does not respond to further operations. This value is only used
214 * if the LNET_MD_MAX_SIZE option is set.
218 * Specifies the behavior of the memory descriptor. A bitwise OR
219 * of the following values can be used:
220 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
221 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
222 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
223 * region is provided by the incoming request. By default, the
224 * offset is maintained locally. When maintained locally, the
225 * offset is incremented by the length of the request so that
226 * the next operation (PUT or GET) will access the next part of
227 * the memory region. Note that only one offset variable exists
228 * per memory descriptor. If both PUT and GET operations are
229 * performed on a memory descriptor, the offset is updated each time.
230 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
231 * be reduced to match the memory available in the region (determined
232 * by subtracting the offset from the length of the memory region).
233 * By default, if the length in the incoming operation is greater
234 * than the amount of memory available, the operation is rejected.
235 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
236 * incoming PUT operations, even if requested. By default,
237 * acknowledgments are sent for PUT operations that request an
238 * acknowledgment. Acknowledgments are never sent for GET operations.
239 * The data sent in the REPLY serves as an implicit acknowledgment.
240 * - LNET_MD_KIOV: The start and length fields specify an array of
242 * - LNET_MD_IOVEC: The start and length fields specify an array of
244 * - LNET_MD_MAX_SIZE: The max_size field is valid.
247 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
248 * capability for memory descriptors. They can't be both set.
249 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
250 * region (i.e. sum of all fragment lengths) must not be less than
253 unsigned int options;
255 * A user-specified value that is associated with the memory
256 * descriptor. The value does not need to be a pointer, but must fit
257 * in the space used by a pointer. This value is recorded in events
258 * associated with operations on this MD.
262 * A handle for the event queue used to log the operations performed on
263 * the memory region. If this argument is a NULL handle (i.e. nullified
264 * by LNetInvalidateHandle()), operations performed on this memory
265 * descriptor are not logged.
267 lnet_handle_eq_t eq_handle;
270 /* Max Transfer Unit (minimum supported everywhere) */
271 #define LNET_MTU_BITS 20
272 #define LNET_MTU (1<<LNET_MTU_BITS)
274 /** limit on the number of fragments in discontiguous MDs */
275 #define LNET_MAX_IOV 256
277 /* Max payload size */
278 #ifndef LNET_MAX_PAYLOAD
279 # error "LNET_MAX_PAYLOAD must be defined in config.h"
281 # if (LNET_MAX_PAYLOAD < LNET_MTU)
282 # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb"
283 # elif defined(__KERNEL__)
284 # if (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV))
285 /* PAGE_SIZE is a constant: check with cpp! */
286 # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb"
292 * Options for the MD structure. See lnet_md_t::options.
294 #define LNET_MD_OP_PUT (1 << 0)
295 /** See lnet_md_t::options. */
296 #define LNET_MD_OP_GET (1 << 1)
297 /** See lnet_md_t::options. */
298 #define LNET_MD_MANAGE_REMOTE (1 << 2)
299 /* unused (1 << 3) */
300 /** See lnet_md_t::options. */
301 #define LNET_MD_TRUNCATE (1 << 4)
302 /** See lnet_md_t::options. */
303 #define LNET_MD_ACK_DISABLE (1 << 5)
304 /** See lnet_md_t::options. */
305 #define LNET_MD_IOVEC (1 << 6)
306 /** See lnet_md_t::options. */
307 #define LNET_MD_MAX_SIZE (1 << 7)
308 /** See lnet_md_t::options. */
309 #define LNET_MD_KIOV (1 << 8)
311 /* For compatibility with Cray Portals */
312 #define LNET_MD_PHYS 0
314 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
315 #define LNET_MD_THRESH_INF (-1)
317 /* NB lustre portals uses struct iovec internally! */
318 typedef struct iovec lnet_md_iovec_t;
321 * A page-based fragment of a MD.
324 /** Pointer to the page where the fragment resides */
325 cfs_page_t *kiov_page;
326 /** Length in bytes of the fragment */
327 unsigned int kiov_len;
329 * Starting offset of the fragment within the page. Note that the
330 * end of the fragment must not pass the end of the page; i.e.,
331 * kiov_len + kiov_offset <= CFS_PAGE_SIZE.
333 unsigned int kiov_offset;
337 /** \addtogroup lnet_eq
341 * Six types of events can be logged in an event queue.
344 /** An incoming GET operation has completed on the MD. */
347 * An incoming PUT operation has completed on the MD. The
348 * underlying layers will not alter the memory (on behalf of this
349 * operation) once this event has been logged.
353 * A REPLY operation has completed. This event is logged after the
354 * data (if any) from the REPLY has been written into the MD.
357 /** An acknowledgment has been received. */
360 * An outgoing send (PUT or GET) operation has completed. This event
361 * is logged after the entire buffer has been sent and it is safe for
362 * the caller to reuse the buffer.
365 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
366 * happen even when the message has not yet been put out on wire.
367 * - It's unsafe to assume that in an outgoing GET operation
368 * the LNET_EVENT_SEND event would happen before the
369 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
370 * LNET_EVENT_ACK events in an outgoing PUT operation.
374 * A MD has been unlinked. Note that LNetMDUnlink() does not
375 * necessarily trigger an LNET_EVENT_UNLINK event.
381 #define LNET_SEQ_BASETYPE long
382 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
383 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
386 * cygwin need the pragma line, not clear if it's needed in other places.
390 #pragma pack(push, 4)
394 * Information about an event on a MD.
397 /** The identifier (nid, pid) of the target. */
398 lnet_process_id_t target;
399 /** The identifier (nid, pid) of the initiator. */
400 lnet_process_id_t initiator;
402 * The NID of the immediate sender. If the request has been forwarded
403 * by routers, this is the NID of the last hop; otherwise it's the
404 * same as the initiator.
407 /** Indicates the type of the event. */
408 lnet_event_kind_t type;
409 /** The portal table index specified in the request */
410 unsigned int pt_index;
411 /** A copy of the match bits specified in the request. */
413 /** The length (in bytes) specified in the request. */
414 unsigned int rlength;
416 * The length (in bytes) of the data that was manipulated by the
417 * operation. For truncated operations, the manipulated length will be
418 * the number of bytes specified by the MD (possibly with an offset,
419 * see lnet_md_t). For all other operations, the manipulated length
420 * will be the length of the requested operation, i.e. rlength.
422 unsigned int mlength;
424 * The handle to the MD associated with the event. The handle may be
425 * invalid if the MD has been unlinked.
427 lnet_handle_md_t md_handle;
429 * A snapshot of the state of the MD immediately after the event has
430 * been processed. In particular, the threshold field in md will
431 * reflect the value of the threshold after the operation occurred.
435 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
440 * Indicates the completion status of the operation. It's 0 for
441 * successful operations, otherwise it's an error code.
445 * Indicates whether the MD has been unlinked. Note that:
446 * - An event with unlinked set is the last event on the MD.
447 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
452 * The displacement (in bytes) into the memory region that the
453 * operation used. The offset can be determined by the operation for
454 * a remote managed MD or by the local MD.
455 * \see lnet_md_t::options
459 * The sequence number for this event. Sequence numbers are unique
462 volatile lnet_seq_t sequence;
469 * Event queue handler function type.
471 * The EQ handler runs for each event that is deposited into the EQ. The
472 * handler is supplied with a pointer to the event that triggered the
473 * handler invocation.
475 * The handler must not block, must be reentrant, and must not call any LNet
476 * API functions. It should return as quickly as possible.
478 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
479 #define LNET_EQ_HANDLER_NONE NULL
482 /** \addtogroup lnet_data
486 * Specify whether an acknowledgment should be sent by target when the PUT
487 * operation completes (i.e., when the data has been written to a MD of the
490 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
491 * acknowledgments can be disabled for a MD.
494 /** Request an acknowledgment */
496 /** Request that no acknowledgment should be generated. */