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 #define LNET_VERSION "0.7.0"
50 /** Portal reserved for LNet's own use.
51 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
53 #define LNET_RESERVED_PORTAL 0
56 * Address of an end-point in an LNet network.
58 * A node can have multiple end-points and hence multiple addresses.
59 * An LNet network can be a simple network (e.g. tcp0) or a network of
60 * LNet networks connected by LNet routers. Therefore an end-point address
61 * has two parts: network ID, and address within a network.
63 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
65 typedef __u64 lnet_nid_t;
67 * ID of a process in a node. Shortened as PID to distinguish from
68 * lnet_process_id_t, the global process ID.
70 typedef __u32 lnet_pid_t;
72 /** wildcard NID that matches any end-point address */
73 #define LNET_NID_ANY ((lnet_nid_t) -1)
74 /** wildcard PID that matches any lnet_pid_t */
75 #define LNET_PID_ANY ((lnet_pid_t) -1)
77 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
78 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
79 #define LNET_PID_LUSTRE 12345
81 #define LNET_TIME_FOREVER (-1)
83 /* how an LNET NID encodes net:address */
84 /** extract the address part of an lnet_nid_t */
86 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
88 return nid & 0xffffffff;
91 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
93 return (nid >> 32) & 0xffffffff;
96 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
98 return (((__u64)net) << 32) | addr;
101 static inline __u32 LNET_NETNUM(__u32 net)
106 static inline __u32 LNET_NETTYP(__u32 net)
108 return (net >> 16) & 0xffff;
111 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
113 return (type << 16) | num;
116 #define WIRE_ATTR __attribute__((packed))
118 /* Packed version of lnet_process_id_t to transfer via network */
121 lnet_pid_t pid; /* node id / process id */
122 } WIRE_ATTR lnet_process_id_packed_t;
124 /* The wire handle's interface cookie only matches one network interface in
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
127 * during that object's lifetime (i.e. no cookie re-use). */
129 __u64 wh_interface_cookie;
130 __u64 wh_object_cookie;
131 } WIRE_ATTR lnet_handle_wire_t;
141 /* The variant fields of the portals message header are aligned on an 8
142 * byte boundary in the message header. Note that all types used in these
143 * wire structs MUST be fixed size and the smaller types are placed at the
145 typedef struct lnet_ack {
146 lnet_handle_wire_t dst_wmd;
149 } WIRE_ATTR lnet_ack_t;
151 typedef struct lnet_put {
152 lnet_handle_wire_t ack_wmd;
157 } WIRE_ATTR lnet_put_t;
159 typedef struct lnet_get {
160 lnet_handle_wire_t return_wmd;
165 } WIRE_ATTR lnet_get_t;
167 typedef struct lnet_reply {
168 lnet_handle_wire_t dst_wmd;
169 } WIRE_ATTR lnet_reply_t;
171 typedef struct lnet_hello {
174 } WIRE_ATTR lnet_hello_t;
181 __u32 type; /* lnet_msg_type_t */
182 __u32 payload_length; /* payload data to follow */
183 /*<------__u64 aligned------->*/
191 } WIRE_ATTR lnet_hdr_t;
193 /* A HELLO message contains a magic number and protocol version
194 * code in the header's dest_nid, the peer's NID in the src_nid, and
195 * LNET_MSG_HELLO in the type field. All other common fields are zero
196 * (including payload_size; i.e. no payload).
197 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
198 * running the same protocol and to find out its NID. These LNDs should
199 * exchange HELLO messages when a connection is first established. Individual
200 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
203 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
204 __u16 version_major; /* increment on incompatible change */
205 __u16 version_minor; /* increment on compatible change */
206 } WIRE_ATTR lnet_magicversion_t;
208 /* PROTO MAGIC for LNDs */
209 #define LNET_PROTO_IB_MAGIC 0x0be91b91
210 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
211 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
212 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
213 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
215 /* Placeholder for a future "unified" protocol across all LNDs */
216 /* Current LNDs that receive a request with this magic will respond with a
217 * * * "stub" reply using their current protocol */
218 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
220 #define LNET_PROTO_TCP_VERSION_MAJOR 1
221 #define LNET_PROTO_TCP_VERSION_MINOR 0
223 /* Acceptor connection request */
225 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
226 __u32 acr_version; /* protocol version */
227 __u64 acr_nid; /* target NID */
228 } WIRE_ATTR lnet_acceptor_connreq_t;
230 #define LNET_PROTO_ACCEPTOR_VERSION 1
232 typedef struct lnet_counters {
244 } WIRE_ATTR lnet_counters_t;
246 #define LNET_NI_STATUS_UP 0x15aac0de
247 #define LNET_NI_STATUS_DOWN 0xdeadface
248 #define LNET_NI_STATUS_INVALID 0x00000000
250 #define LNET_MAX_INTERFACES 16
253 * Objects maintained by the LNet are accessed through handles. Handle types
254 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
255 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
256 * 'me' for match entry).
257 * Each type of object is given a unique handle type to enhance type checking.
258 * The type lnet_handle_any_t can be used when a generic handle is needed.
259 * Every handle value can be converted into a value of type lnet_handle_any_t
260 * without loss of information.
266 typedef lnet_handle_any_t lnet_handle_eq_t;
267 typedef lnet_handle_any_t lnet_handle_md_t;
268 typedef lnet_handle_any_t lnet_handle_me_t;
270 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
273 * Invalidate handle \a h.
275 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
277 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
281 * Compare handles \a h1 and \a h2.
283 * \return 1 if handles are equal, 0 if otherwise.
285 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
287 return (h1.cookie == h2.cookie);
291 * Check whether handle \a h is invalid.
293 * \return 1 if handle is invalid, 0 if valid.
295 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
297 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
311 /** \addtogroup lnet_me
315 * Specifies whether the match entry or memory descriptor should be unlinked
316 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
324 * Values of the type lnet_ins_pos_t are used to control where a new match
325 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
326 * entry before the current entry or before the head of the list. The value
327 * LNET_INS_AFTER is used to insert the new entry after the current entry
328 * or after the last item in the list.
331 /** insert ME before current position or head of the list */
333 /** insert ME after current position or tail of the list */
335 /** attach ME at tail of local CPU partition ME list */
341 /** \addtogroup lnet_md
345 * Defines the visible parts of a memory descriptor. Values of this type
346 * are used to initialize memory descriptors.
350 * Specify the memory region associated with the memory descriptor.
351 * If the options field has:
352 * - LNET_MD_KIOV bit set: The start field points to the starting
353 * address of an array of lnet_kiov_t and the length field specifies
354 * the number of entries in the array. The length can't be bigger
355 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
356 * fragments that are not necessarily mapped in virtal memory.
357 * - LNET_MD_IOVEC bit set: The start field points to the starting
358 * address of an array of struct kvec and the length field specifies
359 * the number of entries in the array. The length can't be bigger
360 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
361 * that have virtual addresses.
362 * - Otherwise: The memory region is contiguous. The start field
363 * specifies the starting address for the memory region and the
364 * length field specifies its length.
366 * When the memory region is fragmented, all fragments but the first
367 * one must start on page boundary, and all but the last must end on
373 * Specifies the maximum number of operations that can be performed
374 * on the memory descriptor. An operation is any action that could
375 * possibly generate an event. In the usual case, the threshold value
376 * is decremented for each operation on the MD. When the threshold
377 * drops to zero, the MD becomes inactive and does not respond to
378 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
379 * there is no bound on the number of operations that may be applied
384 * Specifies the largest incoming request that the memory descriptor
385 * should respond to. When the unused portion of a MD (length -
386 * local offset) falls below this value, the MD becomes inactive and
387 * does not respond to further operations. This value is only used
388 * if the LNET_MD_MAX_SIZE option is set.
392 * Specifies the behavior of the memory descriptor. A bitwise OR
393 * of the following values can be used:
394 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
395 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
396 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
397 * region is provided by the incoming request. By default, the
398 * offset is maintained locally. When maintained locally, the
399 * offset is incremented by the length of the request so that
400 * the next operation (PUT or GET) will access the next part of
401 * the memory region. Note that only one offset variable exists
402 * per memory descriptor. If both PUT and GET operations are
403 * performed on a memory descriptor, the offset is updated each time.
404 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
405 * be reduced to match the memory available in the region (determined
406 * by subtracting the offset from the length of the memory region).
407 * By default, if the length in the incoming operation is greater
408 * than the amount of memory available, the operation is rejected.
409 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
410 * incoming PUT operations, even if requested. By default,
411 * acknowledgments are sent for PUT operations that request an
412 * acknowledgment. Acknowledgments are never sent for GET operations.
413 * The data sent in the REPLY serves as an implicit acknowledgment.
414 * - LNET_MD_KIOV: The start and length fields specify an array of
416 * - LNET_MD_IOVEC: The start and length fields specify an array of
418 * - LNET_MD_MAX_SIZE: The max_size field is valid.
421 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
422 * capability for memory descriptors. They can't be both set.
423 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
424 * region (i.e. sum of all fragment lengths) must not be less than
427 unsigned int options;
429 * A user-specified value that is associated with the memory
430 * descriptor. The value does not need to be a pointer, but must fit
431 * in the space used by a pointer. This value is recorded in events
432 * associated with operations on this MD.
436 * A handle for the event queue used to log the operations performed on
437 * the memory region. If this argument is a NULL handle (i.e. nullified
438 * by LNetInvalidateHandle()), operations performed on this memory
439 * descriptor are not logged.
441 lnet_handle_eq_t eq_handle;
444 /* Max Transfer Unit (minimum supported everywhere).
445 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
446 * these limits are system wide and not interface-local. */
447 #define LNET_MTU_BITS 20
448 #define LNET_MTU (1 << LNET_MTU_BITS)
450 /** limit on the number of fragments in discontiguous MDs */
451 #define LNET_MAX_IOV 256
454 * Options for the MD structure. See lnet_md_t::options.
456 #define LNET_MD_OP_PUT (1 << 0)
457 /** See lnet_md_t::options. */
458 #define LNET_MD_OP_GET (1 << 1)
459 /** See lnet_md_t::options. */
460 #define LNET_MD_MANAGE_REMOTE (1 << 2)
461 /* unused (1 << 3) */
462 /** See lnet_md_t::options. */
463 #define LNET_MD_TRUNCATE (1 << 4)
464 /** See lnet_md_t::options. */
465 #define LNET_MD_ACK_DISABLE (1 << 5)
466 /** See lnet_md_t::options. */
467 #define LNET_MD_IOVEC (1 << 6)
468 /** See lnet_md_t::options. */
469 #define LNET_MD_MAX_SIZE (1 << 7)
470 /** See lnet_md_t::options. */
471 #define LNET_MD_KIOV (1 << 8)
473 /* For compatibility with Cray Portals */
474 #define LNET_MD_PHYS 0
476 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
477 #define LNET_MD_THRESH_INF (-1)
479 /* NB lustre portals uses struct iovec internally! */
480 typedef struct iovec lnet_md_iovec_t;
483 * A page-based fragment of a MD.
486 /** Pointer to the page where the fragment resides */
487 struct page *kiov_page;
488 /** Length in bytes of the fragment */
489 unsigned int kiov_len;
491 * Starting offset of the fragment within the page. Note that the
492 * end of the fragment must not pass the end of the page; i.e.,
493 * kiov_len + kiov_offset <= PAGE_CACHE_SIZE.
495 unsigned int kiov_offset;
499 /** \addtogroup lnet_eq
503 * Six types of events can be logged in an event queue.
506 /** An incoming GET operation has completed on the MD. */
509 * An incoming PUT operation has completed on the MD. The
510 * underlying layers will not alter the memory (on behalf of this
511 * operation) once this event has been logged.
515 * A REPLY operation has completed. This event is logged after the
516 * data (if any) from the REPLY has been written into the MD.
519 /** An acknowledgment has been received. */
522 * An outgoing send (PUT or GET) operation has completed. This event
523 * is logged after the entire buffer has been sent and it is safe for
524 * the caller to reuse the buffer.
527 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
528 * happen even when the message has not yet been put out on wire.
529 * - It's unsafe to assume that in an outgoing GET operation
530 * the LNET_EVENT_SEND event would happen before the
531 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
532 * LNET_EVENT_ACK events in an outgoing PUT operation.
536 * A MD has been unlinked. Note that LNetMDUnlink() does not
537 * necessarily trigger an LNET_EVENT_UNLINK event.
543 #define LNET_SEQ_BASETYPE long
544 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
545 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
548 * Information about an event on a MD.
551 /** The identifier (nid, pid) of the target. */
552 lnet_process_id_t target;
553 /** The identifier (nid, pid) of the initiator. */
554 lnet_process_id_t initiator;
556 * The NID of the immediate sender. If the request has been forwarded
557 * by routers, this is the NID of the last hop; otherwise it's the
558 * same as the initiator.
561 /** Indicates the type of the event. */
562 lnet_event_kind_t type;
563 /** The portal table index specified in the request */
564 unsigned int pt_index;
565 /** A copy of the match bits specified in the request. */
567 /** The length (in bytes) specified in the request. */
568 unsigned int rlength;
570 * The length (in bytes) of the data that was manipulated by the
571 * operation. For truncated operations, the manipulated length will be
572 * the number of bytes specified by the MD (possibly with an offset,
573 * see lnet_md_t). For all other operations, the manipulated length
574 * will be the length of the requested operation, i.e. rlength.
576 unsigned int mlength;
578 * The handle to the MD associated with the event. The handle may be
579 * invalid if the MD has been unlinked.
581 lnet_handle_md_t md_handle;
583 * A snapshot of the state of the MD immediately after the event has
584 * been processed. In particular, the threshold field in md will
585 * reflect the value of the threshold after the operation occurred.
589 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
594 * Indicates the completion status of the operation. It's 0 for
595 * successful operations, otherwise it's an error code.
599 * Indicates whether the MD has been unlinked. Note that:
600 * - An event with unlinked set is the last event on the MD.
601 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
606 * The displacement (in bytes) into the memory region that the
607 * operation used. The offset can be determined by the operation for
608 * a remote managed MD or by the local MD.
609 * \see lnet_md_t::options
613 * The sequence number for this event. Sequence numbers are unique
616 volatile lnet_seq_t sequence;
620 * Event queue handler function type.
622 * The EQ handler runs for each event that is deposited into the EQ. The
623 * handler is supplied with a pointer to the event that triggered the
624 * handler invocation.
626 * The handler must not block, must be reentrant, and must not call any LNet
627 * API functions. It should return as quickly as possible.
629 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
630 #define LNET_EQ_HANDLER_NONE NULL
633 /** \addtogroup lnet_data
637 * Specify whether an acknowledgment should be sent by target when the PUT
638 * operation completes (i.e., when the data has been written to a MD of the
641 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
642 * acknowledgments can be disabled for a MD.
645 /** Request an acknowledgment */
647 /** Request that no acknowledgment should be generated. */