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 <linux/types.h>
44 /** \addtogroup lnet_addr
47 #define LNET_VERSION "0.7.0"
49 /** Portal reserved for LNet's own use.
50 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
52 #define LNET_RESERVED_PORTAL 0
55 * Address of an end-point in an LNet network.
57 * A node can have multiple end-points and hence multiple addresses.
58 * An LNet network can be a simple network (e.g. tcp0) or a network of
59 * LNet networks connected by LNet routers. Therefore an end-point address
60 * has two parts: network ID, and address within a network.
62 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
64 typedef __u64 lnet_nid_t;
66 * ID of a process in a node. Shortened as PID to distinguish from
67 * lnet_process_id_t, the global process ID.
69 typedef __u32 lnet_pid_t;
71 /** wildcard NID that matches any end-point address */
72 #define LNET_NID_ANY ((lnet_nid_t) -1)
73 /** wildcard PID that matches any lnet_pid_t */
74 #define LNET_PID_ANY ((lnet_pid_t) -1)
76 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
77 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
78 #define LNET_PID_LUSTRE 12345
80 #define LNET_TIME_FOREVER (-1)
82 /* how an LNET NID encodes net:address */
83 /** extract the address part of an lnet_nid_t */
85 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
87 return nid & 0xffffffff;
90 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
92 return (nid >> 32) & 0xffffffff;
95 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
97 return (((__u64)net) << 32) | addr;
100 static inline __u32 LNET_NETNUM(__u32 net)
105 static inline __u32 LNET_NETTYP(__u32 net)
107 return (net >> 16) & 0xffff;
110 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
112 return (type << 16) | num;
115 #define WIRE_ATTR __attribute__((packed))
117 /* Packed version of lnet_process_id_t to transfer via network */
120 lnet_pid_t pid; /* node id / process id */
121 } WIRE_ATTR lnet_process_id_packed_t;
123 /* The wire handle's interface cookie only matches one network interface in
124 * one epoch (i.e. new cookie when the interface restarts or the node
125 * reboots). The object cookie only matches one object on that interface
126 * during that object's lifetime (i.e. no cookie re-use). */
128 __u64 wh_interface_cookie;
129 __u64 wh_object_cookie;
130 } WIRE_ATTR lnet_handle_wire_t;
140 /* The variant fields of the portals message header are aligned on an 8
141 * byte boundary in the message header. Note that all types used in these
142 * wire structs MUST be fixed size and the smaller types are placed at the
144 typedef struct lnet_ack {
145 lnet_handle_wire_t dst_wmd;
148 } WIRE_ATTR lnet_ack_t;
150 typedef struct lnet_put {
151 lnet_handle_wire_t ack_wmd;
156 } WIRE_ATTR lnet_put_t;
158 typedef struct lnet_get {
159 lnet_handle_wire_t return_wmd;
164 } WIRE_ATTR lnet_get_t;
166 typedef struct lnet_reply {
167 lnet_handle_wire_t dst_wmd;
168 } WIRE_ATTR lnet_reply_t;
170 typedef struct lnet_hello {
173 } WIRE_ATTR lnet_hello_t;
180 __u32 type; /* lnet_msg_type_t */
181 __u32 payload_length; /* payload data to follow */
182 /*<------__u64 aligned------->*/
190 } WIRE_ATTR lnet_hdr_t;
192 /* A HELLO message contains a magic number and protocol version
193 * code in the header's dest_nid, the peer's NID in the src_nid, and
194 * LNET_MSG_HELLO in the type field. All other common fields are zero
195 * (including payload_size; i.e. no payload).
196 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
197 * running the same protocol and to find out its NID. These LNDs should
198 * exchange HELLO messages when a connection is first established. Individual
199 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
202 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
203 __u16 version_major; /* increment on incompatible change */
204 __u16 version_minor; /* increment on compatible change */
205 } WIRE_ATTR lnet_magicversion_t;
207 /* PROTO MAGIC for LNDs */
208 #define LNET_PROTO_IB_MAGIC 0x0be91b91
209 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
210 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
211 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
212 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
214 /* Placeholder for a future "unified" protocol across all LNDs */
215 /* Current LNDs that receive a request with this magic will respond
216 * with a "stub" reply using their current protocol */
217 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
219 #define LNET_PROTO_TCP_VERSION_MAJOR 1
220 #define LNET_PROTO_TCP_VERSION_MINOR 0
222 /* Acceptor connection request */
224 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
225 __u32 acr_version; /* protocol version */
226 __u64 acr_nid; /* target NID */
227 } WIRE_ATTR lnet_acceptor_connreq_t;
229 #define LNET_PROTO_ACCEPTOR_VERSION 1
231 typedef struct lnet_counters {
243 } WIRE_ATTR lnet_counters_t;
245 #define LNET_NI_STATUS_UP 0x15aac0de
246 #define LNET_NI_STATUS_DOWN 0xdeadface
247 #define LNET_NI_STATUS_INVALID 0x00000000
249 #define LNET_MAX_INTERFACES 16
252 * Objects maintained by the LNet are accessed through handles. Handle types
253 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
254 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
255 * 'me' for match entry).
256 * Each type of object is given a unique handle type to enhance type checking.
257 * The type lnet_handle_any_t can be used when a generic handle is needed.
258 * Every handle value can be converted into a value of type lnet_handle_any_t
259 * without loss of information.
265 typedef lnet_handle_any_t lnet_handle_eq_t;
266 typedef lnet_handle_any_t lnet_handle_md_t;
267 typedef lnet_handle_any_t lnet_handle_me_t;
269 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
272 * Invalidate handle \a h.
274 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
276 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
280 * Compare handles \a h1 and \a h2.
282 * \return 1 if handles are equal, 0 if otherwise.
284 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
286 return (h1.cookie == h2.cookie);
290 * Check whether handle \a h is invalid.
292 * \return 1 if handle is invalid, 0 if valid.
294 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
296 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
310 /** \addtogroup lnet_me
314 * Specifies whether the match entry or memory descriptor should be unlinked
315 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
323 * Values of the type lnet_ins_pos_t are used to control where a new match
324 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
325 * entry before the current entry or before the head of the list. The value
326 * LNET_INS_AFTER is used to insert the new entry after the current entry
327 * or after the last item in the list.
330 /** insert ME before current position or head of the list */
332 /** insert ME after current position or tail of the list */
334 /** attach ME at tail of local CPU partition ME list */
340 /** \addtogroup lnet_md
344 * Defines the visible parts of a memory descriptor. Values of this type
345 * are used to initialize memory descriptors.
349 * Specify the memory region associated with the memory descriptor.
350 * If the options field has:
351 * - LNET_MD_KIOV bit set: The start field points to the starting
352 * address of an array of lnet_kiov_t and the length field specifies
353 * the number of entries in the array. The length can't be bigger
354 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
355 * fragments that are not necessarily mapped in virtal memory.
356 * - LNET_MD_IOVEC bit set: The start field points to the starting
357 * address of an array of struct kvec and the length field specifies
358 * the number of entries in the array. The length can't be bigger
359 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
360 * that have virtual addresses.
361 * - Otherwise: The memory region is contiguous. The start field
362 * specifies the starting address for the memory region and the
363 * length field specifies its length.
365 * When the memory region is fragmented, all fragments but the first
366 * one must start on page boundary, and all but the last must end on
372 * Specifies the maximum number of operations that can be performed
373 * on the memory descriptor. An operation is any action that could
374 * possibly generate an event. In the usual case, the threshold value
375 * is decremented for each operation on the MD. When the threshold
376 * drops to zero, the MD becomes inactive and does not respond to
377 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
378 * there is no bound on the number of operations that may be applied
383 * Specifies the largest incoming request that the memory descriptor
384 * should respond to. When the unused portion of a MD (length -
385 * local offset) falls below this value, the MD becomes inactive and
386 * does not respond to further operations. This value is only used
387 * if the LNET_MD_MAX_SIZE option is set.
391 * Specifies the behavior of the memory descriptor. A bitwise OR
392 * of the following values can be used:
393 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
394 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
395 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
396 * region is provided by the incoming request. By default, the
397 * offset is maintained locally. When maintained locally, the
398 * offset is incremented by the length of the request so that
399 * the next operation (PUT or GET) will access the next part of
400 * the memory region. Note that only one offset variable exists
401 * per memory descriptor. If both PUT and GET operations are
402 * performed on a memory descriptor, the offset is updated each time.
403 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
404 * be reduced to match the memory available in the region (determined
405 * by subtracting the offset from the length of the memory region).
406 * By default, if the length in the incoming operation is greater
407 * than the amount of memory available, the operation is rejected.
408 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
409 * incoming PUT operations, even if requested. By default,
410 * acknowledgments are sent for PUT operations that request an
411 * acknowledgment. Acknowledgments are never sent for GET operations.
412 * The data sent in the REPLY serves as an implicit acknowledgment.
413 * - LNET_MD_KIOV: The start and length fields specify an array of
415 * - LNET_MD_IOVEC: The start and length fields specify an array of
417 * - LNET_MD_MAX_SIZE: The max_size field is valid.
420 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
421 * capability for memory descriptors. They can't be both set.
422 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
423 * region (i.e. sum of all fragment lengths) must not be less than
426 unsigned int options;
428 * A user-specified value that is associated with the memory
429 * descriptor. The value does not need to be a pointer, but must fit
430 * in the space used by a pointer. This value is recorded in events
431 * associated with operations on this MD.
435 * A handle for the event queue used to log the operations performed on
436 * the memory region. If this argument is a NULL handle (i.e. nullified
437 * by LNetInvalidateHandle()), operations performed on this memory
438 * descriptor are not logged.
440 lnet_handle_eq_t eq_handle;
443 /* Max Transfer Unit (minimum supported everywhere).
444 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
445 * these limits are system wide and not interface-local. */
446 #define LNET_MTU_BITS 20
447 #define LNET_MTU (1 << LNET_MTU_BITS)
449 /** limit on the number of fragments in discontiguous MDs */
450 #define LNET_MAX_IOV 256
453 * Options for the MD structure. See lnet_md_t::options.
455 #define LNET_MD_OP_PUT (1 << 0)
456 /** See lnet_md_t::options. */
457 #define LNET_MD_OP_GET (1 << 1)
458 /** See lnet_md_t::options. */
459 #define LNET_MD_MANAGE_REMOTE (1 << 2)
460 /* unused (1 << 3) */
461 /** See lnet_md_t::options. */
462 #define LNET_MD_TRUNCATE (1 << 4)
463 /** See lnet_md_t::options. */
464 #define LNET_MD_ACK_DISABLE (1 << 5)
465 /** See lnet_md_t::options. */
466 #define LNET_MD_IOVEC (1 << 6)
467 /** See lnet_md_t::options. */
468 #define LNET_MD_MAX_SIZE (1 << 7)
469 /** See lnet_md_t::options. */
470 #define LNET_MD_KIOV (1 << 8)
472 /* For compatibility with Cray Portals */
473 #define LNET_MD_PHYS 0
475 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
476 #define LNET_MD_THRESH_INF (-1)
478 /* NB lustre portals uses struct iovec internally! */
479 typedef struct iovec lnet_md_iovec_t;
482 * A page-based fragment of a MD.
485 /** Pointer to the page where the fragment resides */
486 struct page *kiov_page;
487 /** Length in bytes of the fragment */
488 unsigned int kiov_len;
490 * Starting offset of the fragment within the page. Note that the
491 * end of the fragment must not pass the end of the page; i.e.,
492 * kiov_len + kiov_offset <= PAGE_SIZE.
494 unsigned int kiov_offset;
498 /** \addtogroup lnet_eq
502 * Six types of events can be logged in an event queue.
505 /** An incoming GET operation has completed on the MD. */
508 * An incoming PUT operation has completed on the MD. The
509 * underlying layers will not alter the memory (on behalf of this
510 * operation) once this event has been logged.
514 * A REPLY operation has completed. This event is logged after the
515 * data (if any) from the REPLY has been written into the MD.
518 /** An acknowledgment has been received. */
521 * An outgoing send (PUT or GET) operation has completed. This event
522 * is logged after the entire buffer has been sent and it is safe for
523 * the caller to reuse the buffer.
526 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
527 * happen even when the message has not yet been put out on wire.
528 * - It's unsafe to assume that in an outgoing GET operation
529 * the LNET_EVENT_SEND event would happen before the
530 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
531 * LNET_EVENT_ACK events in an outgoing PUT operation.
535 * A MD has been unlinked. Note that LNetMDUnlink() does not
536 * necessarily trigger an LNET_EVENT_UNLINK event.
542 #define LNET_SEQ_BASETYPE long
543 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
544 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
547 * Information about an event on a MD.
550 /** The identifier (nid, pid) of the target. */
551 lnet_process_id_t target;
552 /** The identifier (nid, pid) of the initiator. */
553 lnet_process_id_t initiator;
555 * The NID of the immediate sender. If the request has been forwarded
556 * by routers, this is the NID of the last hop; otherwise it's the
557 * same as the initiator.
560 /** Indicates the type of the event. */
561 lnet_event_kind_t type;
562 /** The portal table index specified in the request */
563 unsigned int pt_index;
564 /** A copy of the match bits specified in the request. */
566 /** The length (in bytes) specified in the request. */
567 unsigned int rlength;
569 * The length (in bytes) of the data that was manipulated by the
570 * operation. For truncated operations, the manipulated length will be
571 * the number of bytes specified by the MD (possibly with an offset,
572 * see lnet_md_t). For all other operations, the manipulated length
573 * will be the length of the requested operation, i.e. rlength.
575 unsigned int mlength;
577 * The handle to the MD associated with the event. The handle may be
578 * invalid if the MD has been unlinked.
580 lnet_handle_md_t md_handle;
582 * A snapshot of the state of the MD immediately after the event has
583 * been processed. In particular, the threshold field in md will
584 * reflect the value of the threshold after the operation occurred.
588 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
593 * Indicates the completion status of the operation. It's 0 for
594 * successful operations, otherwise it's an error code.
598 * Indicates whether the MD has been unlinked. Note that:
599 * - An event with unlinked set is the last event on the MD.
600 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
605 * The displacement (in bytes) into the memory region that the
606 * operation used. The offset can be determined by the operation for
607 * a remote managed MD or by the local MD.
608 * \see lnet_md_t::options
612 * The sequence number for this event. Sequence numbers are unique
615 volatile lnet_seq_t sequence;
619 * Event queue handler function type.
621 * The EQ handler runs for each event that is deposited into the EQ. The
622 * handler is supplied with a pointer to the event that triggered the
623 * handler invocation.
625 * The handler must not block, must be reentrant, and must not call any LNet
626 * API functions. It should return as quickly as possible.
628 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
629 #define LNET_EQ_HANDLER_NONE NULL
632 /** \addtogroup lnet_data
636 * Specify whether an acknowledgment should be sent by target when the PUT
637 * operation completes (i.e., when the data has been written to a MD of the
640 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
641 * acknowledgments can be disabled for a MD.
644 /** Request an acknowledgment */
646 /** Request that no acknowledgment should be generated. */