1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #ifndef __LNET_TYPES_H__
38 #define __LNET_TYPES_H__
43 #include <libcfs/libcfs.h>
45 /** \addtogroup lnet_addr
48 /** Portal reserved for LNet's own use.
49 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
51 #define LNET_RESERVED_PORTAL 0
54 * Address of an end-point in an LNet network.
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.
61 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
63 typedef __u64 lnet_nid_t;
65 * ID of a process in a node. Shortened as PID to distinguish from
66 * lnet_process_id_t, the global process ID.
68 typedef __u32 lnet_pid_t;
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)
76 typedef __u32 lnet_uid_t;
77 #define LNET_UID_ANY ((lnet_uid_t) -1)
80 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
81 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
83 #define LNET_TIME_FOREVER (-1)
86 * Objects maintained by the LNet are accessed through handles. Handle types
87 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
88 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
89 * 'me' for match entry).
90 * Each type of object is given a unique handle type to enhance type checking.
91 * The type lnet_handle_any_t can be used when a generic handle is needed.
92 * Every handle value can be converted into a value of type lnet_handle_any_t
93 * without loss of information.
99 typedef lnet_handle_any_t lnet_handle_eq_t;
100 typedef lnet_handle_any_t lnet_handle_md_t;
101 typedef lnet_handle_any_t lnet_handle_me_t;
103 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
106 * Invalidate handle \a h.
108 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
110 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
114 * Compare handles \a h1 and \a h2.
116 * \return 1 if handles are equal, 0 if otherwise.
118 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
120 return (h1.cookie == h2.cookie);
124 * Check whether handle \a h is invalid.
126 * \return 1 if handle is invalid, 0 if valid.
128 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
130 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
144 /** \addtogroup lnet_me
148 * Specifies whether the match entry or memory descriptor should be unlinked
149 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
157 * Values of the type lnet_ins_pos_t are used to control where a new match
158 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
159 * entry before the current entry or before the head of the list. The value
160 * LNET_INS_AFTER is used to insert the new entry after the current entry
161 * or after the last item in the list.
170 /** \addtogroup lnet_md
174 * Defines the visible parts of a memory descriptor. Values of this type
175 * are used to initialize memory descriptors.
179 * Specify the memory region associated with the memory descriptor.
180 * If the options field has:
181 * - LNET_MD_KIOV bit set: The start field points to the starting
182 * address of an array of lnet_kiov_t and the length field specifies
183 * the number of entries in the array. The length can't be bigger
184 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
185 * fragments that are not necessarily mapped in virtal memory.
186 * - LNET_MD_IOVEC bit set: The start field points to the starting
187 * address of an array of struct iovec and the length field specifies
188 * the number of entries in the array. The length can't be bigger
189 * than LNET_MAX_IOV. The struct iovec is used to describe fragments
190 * that have virtual addresses.
191 * - Otherwise: The memory region is contiguous. The start field
192 * specifies the starting address for the memory region and the
193 * length field specifies its length.
195 * When the memory region is fragmented, all fragments but the first
196 * one must start on page boundary, and all but the last must end on
202 * Specifies the maximum number of operations that can be performed
203 * on the memory descriptor. An operation is any action that could
204 * possibly generate an event. In the usual case, the threshold value
205 * is decremented for each operation on the MD. When the threshold
206 * drops to zero, the MD becomes inactive and does not respond to
207 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
208 * there is no bound on the number of operations that may be applied
213 * Specifies the largest incoming request that the memory descriptor
214 * should respond to. When the unused portion of a MD (length -
215 * local offset) falls below this value, the MD becomes inactive and
216 * does not respond to further operations. This value is only used
217 * if the LNET_MD_MAX_SIZE option is set.
221 * Specifies the behavior of the memory descriptor. A bitwise OR
222 * of the following values can be used:
223 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
224 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
225 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
226 * region is provided by the incoming request. By default, the
227 * offset is maintained locally. When maintained locally, the
228 * offset is incremented by the length of the request so that
229 * the next operation (PUT or GET) will access the next part of
230 * the memory region. Note that only one offset variable exists
231 * per memory descriptor. If both PUT and GET operations are
232 * performed on a memory descriptor, the offset is updated each time.
233 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
234 * be reduced to match the memory available in the region (determined
235 * by subtracting the offset from the length of the memory region).
236 * By default, if the length in the incoming operation is greater
237 * than the amount of memory available, the operation is rejected.
238 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
239 * incoming PUT operations, even if requested. By default,
240 * acknowledgments are sent for PUT operations that request an
241 * acknowledgment. Acknowledgments are never sent for GET operations.
242 * The data sent in the REPLY serves as an implicit acknowledgment.
243 * - LNET_MD_KIOV: The start and length fields specify an array of
245 * - LNET_MD_IOVEC: The start and length fields specify an array of
247 * - LNET_MD_MAX_SIZE: The max_size field is valid.
250 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
251 * capability for memory descriptors. They can't be both set.
252 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
253 * region (i.e. sum of all fragment lengths) must not be less than
256 unsigned int options;
258 * A user-specified value that is associated with the memory
259 * descriptor. The value does not need to be a pointer, but must fit
260 * in the space used by a pointer. This value is recorded in events
261 * associated with operations on this MD.
265 * A handle for the event queue used to log the operations performed on
266 * the memory region. If this argument is a NULL handle (i.e. nullified
267 * by LNetInvalidateHandle()), operations performed on this memory
268 * descriptor are not logged.
270 lnet_handle_eq_t eq_handle;
273 /* Max Transfer Unit (minimum supported everywhere) */
274 #define LNET_MTU_BITS 20
275 #define LNET_MTU (1<<LNET_MTU_BITS)
277 /** limit on the number of fragments in discontiguous MDs */
278 #define LNET_MAX_IOV 256
280 /* Max payload size */
281 #ifndef LNET_MAX_PAYLOAD
282 # error "LNET_MAX_PAYLOAD must be defined in config.h"
284 # if (LNET_MAX_PAYLOAD < LNET_MTU)
285 # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb"
286 # elif defined(__KERNEL__)
287 # if (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV))
288 /* PAGE_SIZE is a constant: check with cpp! */
289 # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb"
295 * Options for the MD structure. See lnet_md_t::options.
297 #define LNET_MD_OP_PUT (1 << 0)
298 /** See lnet_md_t::options. */
299 #define LNET_MD_OP_GET (1 << 1)
300 /** See lnet_md_t::options. */
301 #define LNET_MD_MANAGE_REMOTE (1 << 2)
302 /* unused (1 << 3) */
303 /** See lnet_md_t::options. */
304 #define LNET_MD_TRUNCATE (1 << 4)
305 /** See lnet_md_t::options. */
306 #define LNET_MD_ACK_DISABLE (1 << 5)
307 /** See lnet_md_t::options. */
308 #define LNET_MD_IOVEC (1 << 6)
309 /** See lnet_md_t::options. */
310 #define LNET_MD_MAX_SIZE (1 << 7)
311 /** See lnet_md_t::options. */
312 #define LNET_MD_KIOV (1 << 8)
314 /* For compatibility with Cray Portals */
315 #define LNET_MD_PHYS 0
317 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
318 #define LNET_MD_THRESH_INF (-1)
320 /* NB lustre portals uses struct iovec internally! */
321 typedef struct iovec lnet_md_iovec_t;
324 * A page-based fragment of a MD.
327 /** Pointer to the page where the fragment resides */
328 cfs_page_t *kiov_page;
329 /** Length in bytes of the fragment */
330 unsigned int kiov_len;
332 * Starting offset of the fragment within the page. Note that the
333 * end of the fragment must not pass the end of the page; i.e.,
334 * kiov_len + kiov_offset <= CFS_PAGE_SIZE.
336 unsigned int kiov_offset;
340 /** \addtogroup lnet_eq
344 * Six types of events can be logged in an event queue.
347 /** An incoming GET operation has completed on the MD. */
350 * An incoming PUT operation has completed on the MD. The
351 * underlying layers will not alter the memory (on behalf of this
352 * operation) once this event has been logged.
356 * A REPLY operation has completed. This event is logged after the
357 * data (if any) from the REPLY has been written into the MD.
360 /** An acknowledgment has been received. */
363 * An outgoing send (PUT or GET) operation has completed. This event
364 * is logged after the entire buffer has been sent and it is safe for
365 * the caller to reuse the buffer.
368 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
369 * happen even when the message has not yet been put out on wire.
370 * - It's unsafe to assume that in an outgoing GET operation
371 * the LNET_EVENT_SEND event would happen before the
372 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
373 * LNET_EVENT_ACK events in an outgoing PUT operation.
377 * A MD has been unlinked. Note that LNetMDUnlink() does not
378 * necessarily trigger an LNET_EVENT_UNLINK event.
384 #define LNET_SEQ_BASETYPE long
385 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
386 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
389 * cygwin need the pragma line, not clear if it's needed in other places.
393 #pragma pack(push, 4)
397 * Information about an event on a MD.
400 /** The identifier (nid, pid) of the target. */
401 lnet_process_id_t target;
402 /** The identifier (nid, pid) of the initiator. */
403 lnet_process_id_t initiator;
405 * The NID of the immediate sender. If the request has been forwarded
406 * by routers, this is the NID of the last hop; otherwise it's the
407 * same as the initiator.
410 /** Indicates the type of the event. */
411 lnet_event_kind_t type;
412 /** The portal table index specified in the request */
413 unsigned int pt_index;
414 /** A copy of the match bits specified in the request. */
416 /** The length (in bytes) specified in the request. */
417 unsigned int rlength;
419 * The length (in bytes) of the data that was manipulated by the
420 * operation. For truncated operations, the manipulated length will be
421 * the number of bytes specified by the MD (possibly with an offset,
422 * see lnet_md_t). For all other operations, the manipulated length
423 * will be the length of the requested operation, i.e. rlength.
425 unsigned int mlength;
427 * The handle to the MD associated with the event. The handle may be
428 * invalid if the MD has been unlinked.
430 lnet_handle_md_t md_handle;
432 * A snapshot of the state of the MD immediately after the event has
433 * been processed. In particular, the threshold field in md will
434 * reflect the value of the threshold after the operation occurred.
438 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
443 * Indicates the completion status of the operation. It's 0 for
444 * successful operations, otherwise it's an error code.
448 * Indicates whether the MD has been unlinked. Note that:
449 * - An event with unlinked set is the last event on the MD.
450 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
455 * The displacement (in bytes) into the memory region that the
456 * operation used. The offset can be determined by the operation for
457 * a remote managed MD or by the local MD.
458 * \see lnet_md_t::options
466 * The sequence number for this event. Sequence numbers are unique
469 volatile lnet_seq_t sequence;
476 * Event queue handler function type.
478 * The EQ handler runs for each event that is deposited into the EQ. The
479 * handler is supplied with a pointer to the event that triggered the
480 * handler invocation.
482 * The handler must not block, must be reentrant, and must not call any LNet
483 * API functions. It should return as quickly as possible.
485 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
486 #define LNET_EQ_HANDLER_NONE NULL
489 /** \addtogroup lnet_data
493 * Specify whether an acknowledgment should be sent by target when the PUT
494 * operation completes (i.e., when the data has been written to a MD of the
497 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
498 * acknowledgments can be disabled for a MD.
501 /** Request an acknowledgment */
503 /** Request that no acknowledgment should be generated. */