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.
30 * Copyright (c) 2012, 2014, Intel Corporation.
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/types.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)
75 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
76 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
77 #define LNET_PID_LUSTRE 12345
79 #define LNET_TIME_FOREVER (-1)
81 /* how an LNET NID encodes net:address */
82 /** extract the address part of an lnet_nid_t */
84 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
86 return nid & 0xffffffff;
89 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
91 return (nid >> 32) & 0xffffffff;
94 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
96 return (((__u64)net) << 32) | addr;
99 static inline __u32 LNET_NETNUM(__u32 net)
104 static inline __u32 LNET_NETTYP(__u32 net)
106 return (net >> 16) & 0xffff;
109 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
111 return (type << 16) | num;
115 * Objects maintained by the LNet are accessed through handles. Handle types
116 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
117 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
118 * 'me' for match entry).
119 * Each type of object is given a unique handle type to enhance type checking.
120 * The type lnet_handle_any_t can be used when a generic handle is needed.
121 * Every handle value can be converted into a value of type lnet_handle_any_t
122 * without loss of information.
128 typedef lnet_handle_any_t lnet_handle_eq_t;
129 typedef lnet_handle_any_t lnet_handle_md_t;
130 typedef lnet_handle_any_t lnet_handle_me_t;
132 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
135 * Invalidate handle \a h.
137 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
139 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
143 * Compare handles \a h1 and \a h2.
145 * \return 1 if handles are equal, 0 if otherwise.
147 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
149 return (h1.cookie == h2.cookie);
153 * Check whether handle \a h is invalid.
155 * \return 1 if handle is invalid, 0 if valid.
157 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
159 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
173 /** \addtogroup lnet_me
177 * Specifies whether the match entry or memory descriptor should be unlinked
178 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
186 * Values of the type lnet_ins_pos_t are used to control where a new match
187 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
188 * entry before the current entry or before the head of the list. The value
189 * LNET_INS_AFTER is used to insert the new entry after the current entry
190 * or after the last item in the list.
193 /** insert ME before current position or head of the list */
195 /** insert ME after current position or tail of the list */
197 /** attach ME at tail of local CPU partition ME list */
203 /** \addtogroup lnet_md
207 * Defines the visible parts of a memory descriptor. Values of this type
208 * are used to initialize memory descriptors.
212 * Specify the memory region associated with the memory descriptor.
213 * If the options field has:
214 * - LNET_MD_KIOV bit set: The start field points to the starting
215 * address of an array of lnet_kiov_t and the length field specifies
216 * the number of entries in the array. The length can't be bigger
217 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
218 * fragments that are not necessarily mapped in virtal memory.
219 * - LNET_MD_IOVEC bit set: The start field points to the starting
220 * address of an array of struct iovec and the length field specifies
221 * the number of entries in the array. The length can't be bigger
222 * than LNET_MAX_IOV. The struct iovec is used to describe fragments
223 * that have virtual addresses.
224 * - Otherwise: The memory region is contiguous. The start field
225 * specifies the starting address for the memory region and the
226 * length field specifies its length.
228 * When the memory region is fragmented, all fragments but the first
229 * one must start on page boundary, and all but the last must end on
235 * Specifies the maximum number of operations that can be performed
236 * on the memory descriptor. An operation is any action that could
237 * possibly generate an event. In the usual case, the threshold value
238 * is decremented for each operation on the MD. When the threshold
239 * drops to zero, the MD becomes inactive and does not respond to
240 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
241 * there is no bound on the number of operations that may be applied
246 * Specifies the largest incoming request that the memory descriptor
247 * should respond to. When the unused portion of a MD (length -
248 * local offset) falls below this value, the MD becomes inactive and
249 * does not respond to further operations. This value is only used
250 * if the LNET_MD_MAX_SIZE option is set.
254 * Specifies the behavior of the memory descriptor. A bitwise OR
255 * of the following values can be used:
256 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
257 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
258 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
259 * region is provided by the incoming request. By default, the
260 * offset is maintained locally. When maintained locally, the
261 * offset is incremented by the length of the request so that
262 * the next operation (PUT or GET) will access the next part of
263 * the memory region. Note that only one offset variable exists
264 * per memory descriptor. If both PUT and GET operations are
265 * performed on a memory descriptor, the offset is updated each time.
266 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
267 * be reduced to match the memory available in the region (determined
268 * by subtracting the offset from the length of the memory region).
269 * By default, if the length in the incoming operation is greater
270 * than the amount of memory available, the operation is rejected.
271 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
272 * incoming PUT operations, even if requested. By default,
273 * acknowledgments are sent for PUT operations that request an
274 * acknowledgment. Acknowledgments are never sent for GET operations.
275 * The data sent in the REPLY serves as an implicit acknowledgment.
276 * - LNET_MD_KIOV: The start and length fields specify an array of
278 * - LNET_MD_IOVEC: The start and length fields specify an array of
280 * - LNET_MD_MAX_SIZE: The max_size field is valid.
283 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
284 * capability for memory descriptors. They can't be both set.
285 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
286 * region (i.e. sum of all fragment lengths) must not be less than
289 unsigned int options;
291 * A user-specified value that is associated with the memory
292 * descriptor. The value does not need to be a pointer, but must fit
293 * in the space used by a pointer. This value is recorded in events
294 * associated with operations on this MD.
298 * A handle for the event queue used to log the operations performed on
299 * the memory region. If this argument is a NULL handle (i.e. nullified
300 * by LNetInvalidateHandle()), operations performed on this memory
301 * descriptor are not logged.
303 lnet_handle_eq_t eq_handle;
306 /* Max Transfer Unit (minimum supported everywhere).
307 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
308 * these limits are system wide and not interface-local. */
309 #define LNET_MTU_BITS 20
310 #define LNET_MTU (1 << LNET_MTU_BITS)
312 /** limit on the number of fragments in discontiguous MDs */
313 #define LNET_MAX_IOV 256
315 /* Max payload size */
316 #ifndef CONFIG_LNET_MAX_PAYLOAD
317 # error "CONFIG_LNET_MAX_PAYLOAD must be defined in config.h"
319 # define LNET_MAX_PAYLOAD CONFIG_LNET_MAX_PAYLOAD
320 # if (LNET_MAX_PAYLOAD < LNET_MTU)
321 # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb"
322 # elif defined(__KERNEL__)
323 # if (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV))
324 /* PAGE_SIZE is a constant: check with cpp! */
325 # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb"
331 * Options for the MD structure. See lnet_md_t::options.
333 #define LNET_MD_OP_PUT (1 << 0)
334 /** See lnet_md_t::options. */
335 #define LNET_MD_OP_GET (1 << 1)
336 /** See lnet_md_t::options. */
337 #define LNET_MD_MANAGE_REMOTE (1 << 2)
338 /* unused (1 << 3) */
339 /** See lnet_md_t::options. */
340 #define LNET_MD_TRUNCATE (1 << 4)
341 /** See lnet_md_t::options. */
342 #define LNET_MD_ACK_DISABLE (1 << 5)
343 /** See lnet_md_t::options. */
344 #define LNET_MD_IOVEC (1 << 6)
345 /** See lnet_md_t::options. */
346 #define LNET_MD_MAX_SIZE (1 << 7)
347 /** See lnet_md_t::options. */
348 #define LNET_MD_KIOV (1 << 8)
350 /* For compatibility with Cray Portals */
351 #define LNET_MD_PHYS 0
353 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
354 #define LNET_MD_THRESH_INF (-1)
356 /* NB lustre portals uses struct iovec internally! */
357 typedef struct iovec lnet_md_iovec_t;
360 * A page-based fragment of a MD.
363 /** Pointer to the page where the fragment resides */
364 struct page *kiov_page;
365 /** Length in bytes of the fragment */
366 unsigned int kiov_len;
368 * Starting offset of the fragment within the page. Note that the
369 * end of the fragment must not pass the end of the page; i.e.,
370 * kiov_len + kiov_offset <= PAGE_CACHE_SIZE.
372 unsigned int kiov_offset;
376 /** \addtogroup lnet_eq
380 * Six types of events can be logged in an event queue.
383 /** An incoming GET operation has completed on the MD. */
386 * An incoming PUT operation has completed on the MD. The
387 * underlying layers will not alter the memory (on behalf of this
388 * operation) once this event has been logged.
392 * A REPLY operation has completed. This event is logged after the
393 * data (if any) from the REPLY has been written into the MD.
396 /** An acknowledgment has been received. */
399 * An outgoing send (PUT or GET) operation has completed. This event
400 * is logged after the entire buffer has been sent and it is safe for
401 * the caller to reuse the buffer.
404 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
405 * happen even when the message has not yet been put out on wire.
406 * - It's unsafe to assume that in an outgoing GET operation
407 * the LNET_EVENT_SEND event would happen before the
408 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
409 * LNET_EVENT_ACK events in an outgoing PUT operation.
413 * A MD has been unlinked. Note that LNetMDUnlink() does not
414 * necessarily trigger an LNET_EVENT_UNLINK event.
420 #define LNET_SEQ_BASETYPE long
421 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
422 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
425 * Information about an event on a MD.
428 /** The identifier (nid, pid) of the target. */
429 lnet_process_id_t target;
430 /** The identifier (nid, pid) of the initiator. */
431 lnet_process_id_t initiator;
433 * The NID of the immediate sender. If the request has been forwarded
434 * by routers, this is the NID of the last hop; otherwise it's the
435 * same as the initiator.
438 /** Indicates the type of the event. */
439 lnet_event_kind_t type;
440 /** The portal table index specified in the request */
441 unsigned int pt_index;
442 /** A copy of the match bits specified in the request. */
444 /** The length (in bytes) specified in the request. */
445 unsigned int rlength;
447 * The length (in bytes) of the data that was manipulated by the
448 * operation. For truncated operations, the manipulated length will be
449 * the number of bytes specified by the MD (possibly with an offset,
450 * see lnet_md_t). For all other operations, the manipulated length
451 * will be the length of the requested operation, i.e. rlength.
453 unsigned int mlength;
455 * The handle to the MD associated with the event. The handle may be
456 * invalid if the MD has been unlinked.
458 lnet_handle_md_t md_handle;
460 * A snapshot of the state of the MD immediately after the event has
461 * been processed. In particular, the threshold field in md will
462 * reflect the value of the threshold after the operation occurred.
466 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
471 * Indicates the completion status of the operation. It's 0 for
472 * successful operations, otherwise it's an error code.
476 * Indicates whether the MD has been unlinked. Note that:
477 * - An event with unlinked set is the last event on the MD.
478 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
483 * The displacement (in bytes) into the memory region that the
484 * operation used. The offset can be determined by the operation for
485 * a remote managed MD or by the local MD.
486 * \see lnet_md_t::options
490 * The sequence number for this event. Sequence numbers are unique
493 volatile lnet_seq_t sequence;
497 * Event queue handler function type.
499 * The EQ handler runs for each event that is deposited into the EQ. The
500 * handler is supplied with a pointer to the event that triggered the
501 * handler invocation.
503 * The handler must not block, must be reentrant, and must not call any LNet
504 * API functions. It should return as quickly as possible.
506 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
507 #define LNET_EQ_HANDLER_NONE NULL
510 /** \addtogroup lnet_data
514 * Specify whether an acknowledgment should be sent by target when the PUT
515 * operation completes (i.e., when the data has been written to a MD of the
518 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
519 * acknowledgments can be disabled for a MD.
522 /** Request an acknowledgment */
524 /** Request that no acknowledgment should be generated. */