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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012 - 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Seagate, Inc.
33 #ifndef __LNET_TYPES_H__
34 #define __LNET_TYPES_H__
36 #include <linux/types.h>
42 #define LNET_VERSION "0.6.0"
44 /** \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;
114 #define WIRE_ATTR __packed
116 /* Packed version of lnet_process_id_t to transfer via network */
118 /* node id / process id */
121 } 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).
130 __u64 wh_interface_cookie;
131 __u64 wh_object_cookie;
132 } WIRE_ATTR lnet_handle_wire_t;
143 * The variant fields of the portals message header are aligned on an 8
144 * byte boundary in the message header. Note that all types used in these
145 * wire structs MUST be fixed size and the smaller types are placed at the
148 typedef struct lnet_ack {
149 lnet_handle_wire_t dst_wmd;
152 } WIRE_ATTR lnet_ack_t;
154 typedef struct lnet_put {
155 lnet_handle_wire_t ack_wmd;
160 } WIRE_ATTR lnet_put_t;
162 typedef struct lnet_get {
163 lnet_handle_wire_t return_wmd;
168 } WIRE_ATTR lnet_get_t;
170 typedef struct lnet_reply {
171 lnet_handle_wire_t dst_wmd;
172 } WIRE_ATTR lnet_reply_t;
174 typedef struct lnet_hello {
177 } WIRE_ATTR lnet_hello_t;
184 __u32 type; /* lnet_msg_type_t */
185 __u32 payload_length; /* payload data to follow */
186 /*<------__u64 aligned------->*/
194 } WIRE_ATTR lnet_hdr_t;
197 * A HELLO message contains a magic number and protocol version
198 * code in the header's dest_nid, the peer's NID in the src_nid, and
199 * LNET_MSG_HELLO in the type field. All other common fields are zero
200 * (including payload_size; i.e. no payload).
201 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
202 * running the same protocol and to find out its NID. These LNDs should
203 * exchange HELLO messages when a connection is first established. Individual
204 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
207 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
208 __u16 version_major; /* increment on incompatible change */
209 __u16 version_minor; /* increment on compatible change */
210 } WIRE_ATTR lnet_magicversion_t;
212 /* PROTO MAGIC for LNDs */
213 #define LNET_PROTO_IB_MAGIC 0x0be91b91
214 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
215 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
216 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
217 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
219 /* Placeholder for a future "unified" protocol across all LNDs */
221 * Current LNDs that receive a request with this magic will respond with a
222 * "stub" reply using their current protocol
224 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
226 #define LNET_PROTO_TCP_VERSION_MAJOR 1
227 #define LNET_PROTO_TCP_VERSION_MINOR 0
229 /* Acceptor connection request */
231 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
232 __u32 acr_version; /* protocol version */
233 __u64 acr_nid; /* target NID */
234 } WIRE_ATTR lnet_acceptor_connreq_t;
236 #define LNET_PROTO_ACCEPTOR_VERSION 1
242 } WIRE_ATTR lnet_ni_status_t;
249 lnet_ni_status_t pi_ni[0];
250 } WIRE_ATTR lnet_ping_info_t;
252 typedef struct lnet_counters {
264 } WIRE_ATTR lnet_counters_t;
266 #define LNET_NI_STATUS_UP 0x15aac0de
267 #define LNET_NI_STATUS_DOWN 0xdeadface
268 #define LNET_NI_STATUS_INVALID 0x00000000
270 #define LNET_MAX_INTERFACES 16
273 * Objects maintained by the LNet are accessed through handles. Handle types
274 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
275 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
276 * 'me' for match entry).
277 * Each type of object is given a unique handle type to enhance type checking.
278 * The type lnet_handle_any_t can be used when a generic handle is needed.
279 * Every handle value can be converted into a value of type lnet_handle_any_t
280 * without loss of information.
286 typedef lnet_handle_any_t lnet_handle_eq_t;
287 typedef lnet_handle_any_t lnet_handle_md_t;
288 typedef lnet_handle_any_t lnet_handle_me_t;
290 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
293 * Invalidate handle \a h.
295 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
297 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
301 * Compare handles \a h1 and \a h2.
303 * \return 1 if handles are equal, 0 if otherwise.
305 static inline int LNetHandleIsEqual(lnet_handle_any_t h1, lnet_handle_any_t h2)
307 return h1.cookie == h2.cookie;
311 * Check whether handle \a h is invalid.
313 * \return 1 if handle is invalid, 0 if valid.
315 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
317 return h.cookie == LNET_WIRE_HANDLE_COOKIE_NONE;
331 /** \addtogroup lnet_me
336 * Specifies whether the match entry or memory descriptor should be unlinked
337 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
345 * Values of the type lnet_ins_pos_t are used to control where a new match
346 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
347 * entry before the current entry or before the head of the list. The value
348 * LNET_INS_AFTER is used to insert the new entry after the current entry
349 * or after the last item in the list.
352 /** insert ME before current position or head of the list */
354 /** insert ME after current position or tail of the list */
356 /** attach ME at tail of local CPU partition ME list */
362 /** \addtogroup lnet_md
367 * Defines the visible parts of a memory descriptor. Values of this type
368 * are used to initialize memory descriptors.
372 * Specify the memory region associated with the memory descriptor.
373 * If the options field has:
374 * - LNET_MD_KIOV bit set: The start field points to the starting
375 * address of an array of lnet_kiov_t and the length field specifies
376 * the number of entries in the array. The length can't be bigger
377 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
378 * fragments that are not necessarily mapped in virtual memory.
379 * - LNET_MD_IOVEC bit set: The start field points to the starting
380 * address of an array of struct iovec and the length field specifies
381 * the number of entries in the array. The length can't be bigger
382 * than LNET_MAX_IOV. The struct iovec is used to describe fragments
383 * that have virtual addresses.
384 * - Otherwise: The memory region is contiguous. The start field
385 * specifies the starting address for the memory region and the
386 * length field specifies its length.
388 * When the memory region is fragmented, all fragments but the first
389 * one must start on page boundary, and all but the last must end on
395 * Specifies the maximum number of operations that can be performed
396 * on the memory descriptor. An operation is any action that could
397 * possibly generate an event. In the usual case, the threshold value
398 * is decremented for each operation on the MD. When the threshold
399 * drops to zero, the MD becomes inactive and does not respond to
400 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
401 * there is no bound on the number of operations that may be applied
406 * Specifies the largest incoming request that the memory descriptor
407 * should respond to. When the unused portion of a MD (length -
408 * local offset) falls below this value, the MD becomes inactive and
409 * does not respond to further operations. This value is only used
410 * if the LNET_MD_MAX_SIZE option is set.
414 * Specifies the behavior of the memory descriptor. A bitwise OR
415 * of the following values can be used:
416 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
417 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
418 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
419 * region is provided by the incoming request. By default, the
420 * offset is maintained locally. When maintained locally, the
421 * offset is incremented by the length of the request so that
422 * the next operation (PUT or GET) will access the next part of
423 * the memory region. Note that only one offset variable exists
424 * per memory descriptor. If both PUT and GET operations are
425 * performed on a memory descriptor, the offset is updated each time.
426 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
427 * be reduced to match the memory available in the region (determined
428 * by subtracting the offset from the length of the memory region).
429 * By default, if the length in the incoming operation is greater
430 * than the amount of memory available, the operation is rejected.
431 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
432 * incoming PUT operations, even if requested. By default,
433 * acknowledgments are sent for PUT operations that request an
434 * acknowledgment. Acknowledgments are never sent for GET operations.
435 * The data sent in the REPLY serves as an implicit acknowledgment.
436 * - LNET_MD_KIOV: The start and length fields specify an array of
438 * - LNET_MD_IOVEC: The start and length fields specify an array of
440 * - LNET_MD_MAX_SIZE: The max_size field is valid.
443 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
444 * capability for memory descriptors. They can't be both set.
445 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
446 * region (i.e. sum of all fragment lengths) must not be less than
449 unsigned int options;
451 * A user-specified value that is associated with the memory
452 * descriptor. The value does not need to be a pointer, but must fit
453 * in the space used by a pointer. This value is recorded in events
454 * associated with operations on this MD.
458 * A handle for the event queue used to log the operations performed on
459 * the memory region. If this argument is a NULL handle (i.e. nullified
460 * by LNetInvalidateHandle()), operations performed on this memory
461 * descriptor are not logged.
463 lnet_handle_eq_t eq_handle;
467 * Max Transfer Unit (minimum supported everywhere).
468 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
469 * these limits are system wide and not interface-local.
471 #define LNET_MTU_BITS 20
472 #define LNET_MTU (1 << LNET_MTU_BITS)
474 /** limit on the number of fragments in discontiguous MDs */
475 #define LNET_MAX_IOV 256
478 * Options for the MD structure. See lnet_md_t::options.
480 #define LNET_MD_OP_PUT (1 << 0)
481 /** See lnet_md_t::options. */
482 #define LNET_MD_OP_GET (1 << 1)
483 /** See lnet_md_t::options. */
484 #define LNET_MD_MANAGE_REMOTE (1 << 2)
485 /* unused (1 << 3) */
486 /** See lnet_md_t::options. */
487 #define LNET_MD_TRUNCATE (1 << 4)
488 /** See lnet_md_t::options. */
489 #define LNET_MD_ACK_DISABLE (1 << 5)
490 /** See lnet_md_t::options. */
491 #define LNET_MD_IOVEC (1 << 6)
492 /** See lnet_md_t::options. */
493 #define LNET_MD_MAX_SIZE (1 << 7)
494 /** See lnet_md_t::options. */
495 #define LNET_MD_KIOV (1 << 8)
497 /* For compatibility with Cray Portals */
498 #define LNET_MD_PHYS 0
500 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
501 #define LNET_MD_THRESH_INF (-1)
503 /* NB lustre portals uses struct iovec internally! */
504 typedef struct iovec lnet_md_iovec_t;
506 typedef struct bio_vec lnet_kiov_t;
509 /** \addtogroup lnet_eq
514 * Six types of events can be logged in an event queue.
517 /** An incoming GET operation has completed on the MD. */
520 * An incoming PUT operation has completed on the MD. The
521 * underlying layers will not alter the memory (on behalf of this
522 * operation) once this event has been logged.
526 * A REPLY operation has completed. This event is logged after the
527 * data (if any) from the REPLY has been written into the MD.
530 /** An acknowledgment has been received. */
533 * An outgoing send (PUT or GET) operation has completed. This event
534 * is logged after the entire buffer has been sent and it is safe for
535 * the caller to reuse the buffer.
538 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
539 * happen even when the message has not yet been put out on wire.
540 * - It's unsafe to assume that in an outgoing GET operation
541 * the LNET_EVENT_SEND event would happen before the
542 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
543 * LNET_EVENT_ACK events in an outgoing PUT operation.
547 * A MD has been unlinked. Note that LNetMDUnlink() does not
548 * necessarily trigger an LNET_EVENT_UNLINK event.
554 #define LNET_SEQ_BASETYPE long
555 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
556 #define LNET_SEQ_GT(a, b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
559 * Information about an event on a MD.
562 /** The identifier (nid, pid) of the target. */
563 lnet_process_id_t target;
564 /** The identifier (nid, pid) of the initiator. */
565 lnet_process_id_t initiator;
567 * The NID of the immediate sender. If the request has been forwarded
568 * by routers, this is the NID of the last hop; otherwise it's the
569 * same as the initiator.
572 /** Indicates the type of the event. */
573 lnet_event_kind_t type;
574 /** The portal table index specified in the request */
575 unsigned int pt_index;
576 /** A copy of the match bits specified in the request. */
578 /** The length (in bytes) specified in the request. */
579 unsigned int rlength;
581 * The length (in bytes) of the data that was manipulated by the
582 * operation. For truncated operations, the manipulated length will be
583 * the number of bytes specified by the MD (possibly with an offset,
584 * see lnet_md_t). For all other operations, the manipulated length
585 * will be the length of the requested operation, i.e. rlength.
587 unsigned int mlength;
589 * The handle to the MD associated with the event. The handle may be
590 * invalid if the MD has been unlinked.
592 lnet_handle_md_t md_handle;
594 * A snapshot of the state of the MD immediately after the event has
595 * been processed. In particular, the threshold field in md will
596 * reflect the value of the threshold after the operation occurred.
600 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
605 * Indicates the completion status of the operation. It's 0 for
606 * successful operations, otherwise it's an error code.
610 * Indicates whether the MD has been unlinked. Note that:
611 * - An event with unlinked set is the last event on the MD.
612 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
617 * The displacement (in bytes) into the memory region that the
618 * operation used. The offset can be determined by the operation for
619 * a remote managed MD or by the local MD.
620 * \see lnet_md_t::options
624 * The sequence number for this event. Sequence numbers are unique
627 volatile lnet_seq_t sequence;
631 * Event queue handler function type.
633 * The EQ handler runs for each event that is deposited into the EQ. The
634 * handler is supplied with a pointer to the event that triggered the
635 * handler invocation.
637 * The handler must not block, must be reentrant, and must not call any LNet
638 * API functions. It should return as quickly as possible.
640 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
641 #define LNET_EQ_HANDLER_NONE NULL
644 /** \addtogroup lnet_data
649 * Specify whether an acknowledgment should be sent by target when the PUT
650 * operation completes (i.e., when the data has been written to a MD of the
653 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
654 * acknowledgments can be disabled for a MD.
657 /** Request an acknowledgment */
659 /** Request that no acknowledgment should be generated. */