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 Sun Microsystems, Inc.
33 #ifndef __UAPI_LNET_TYPES_H__
34 #define __UAPI_LNET_TYPES_H__
39 #include <linux/types.h>
40 /** \addtogroup lnet_addr
43 #define LNET_VERSION "0.7.0"
45 /** Portal reserved for LNet's own use.
46 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
48 #define LNET_RESERVED_PORTAL 0
51 * Address of an end-point in an LNet network.
53 * A node can have multiple end-points and hence multiple addresses.
54 * An LNet network can be a simple network (e.g. tcp0) or a network of
55 * LNet networks connected by LNet routers. Therefore an end-point address
56 * has two parts: network ID, and address within a network.
58 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
60 typedef __u64 lnet_nid_t;
63 * ID of a process in a node. Shortened as PID to distinguish from
64 * lnet_process_id, the global process ID.
66 typedef __u32 lnet_pid_t;
68 /** wildcard NID that matches any end-point address */
69 #define LNET_NID_ANY ((lnet_nid_t) -1)
70 /** wildcard PID that matches any lnet_pid_t */
71 #define LNET_PID_ANY ((lnet_pid_t) -1)
73 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
74 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
75 #define LNET_PID_LUSTRE 12345
77 /* how an LNET NID encodes net:address */
78 /** extract the address part of an lnet_nid_t */
80 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
82 return nid & 0xffffffff;
85 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
87 return (nid >> 32) & 0xffffffff;
90 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
92 return (((__u64)net) << 32) | addr;
95 static inline __u32 LNET_NETNUM(__u32 net)
100 static inline __u32 LNET_NETTYP(__u32 net)
102 return (net >> 16) & 0xffff;
105 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
107 return (type << 16) | num;
110 #define WIRE_ATTR __attribute__((packed))
112 /* Packed version of struct lnet_process_id to transfer via network */
113 typedef struct lnet_process_id_packed {
115 lnet_pid_t pid; /* node id / process id */
116 } WIRE_ATTR lnet_process_id_packed;
118 /* The wire handle's interface cookie only matches one network interface in
119 * one epoch (i.e. new cookie when the interface restarts or the node
120 * reboots). The object cookie only matches one object on that interface
121 * during that object's lifetime (i.e. no cookie re-use). */
122 typedef struct lnet_handle_wire {
123 __u64 wh_interface_cookie;
124 __u64 wh_object_cookie;
125 } WIRE_ATTR lnet_handle_wire_t;
127 typedef enum lnet_msg_type {
135 /* The variant fields of the portals message header are aligned on an 8
136 * byte boundary in the message header. Note that all types used in these
137 * wire structs MUST be fixed size and the smaller types are placed at the
140 struct lnet_handle_wire dst_wmd;
146 struct lnet_handle_wire ack_wmd;
154 struct lnet_handle_wire return_wmd;
162 struct lnet_handle_wire dst_wmd;
170 typedef struct lnet_hdr {
175 __u32 type; /* enum lnet_msg_type */
176 __u32 payload_length; /* payload data to follow */
177 /*<------__u64 aligned------->*/
182 struct lnet_reply reply;
183 struct lnet_hello hello;
185 } WIRE_ATTR lnet_hdr_t;
187 /* A HELLO message contains a magic number and protocol version
188 * code in the header's dest_nid, the peer's NID in the src_nid, and
189 * LNET_MSG_HELLO in the type field. All other common fields are zero
190 * (including payload_size; i.e. no payload).
191 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
192 * running the same protocol and to find out its NID. These LNDs should
193 * exchange HELLO messages when a connection is first established. Individual
194 * LNDs can put whatever else they fancy in lnet_hdr::msg.
196 typedef struct lnet_magicversion {
197 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
198 __u16 version_major; /* increment on incompatible change */
199 __u16 version_minor; /* increment on compatible change */
200 } WIRE_ATTR lnet_magic_version_t;
202 /* PROTO MAGIC for LNDs */
203 #define LNET_PROTO_IB_MAGIC 0x0be91b91
204 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
205 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
206 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
207 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
209 /* Placeholder for a future "unified" protocol across all LNDs */
210 /* Current LNDs that receive a request with this magic will respond
211 * with a "stub" reply using their current protocol */
212 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
214 #define LNET_PROTO_TCP_VERSION_MAJOR 1
215 #define LNET_PROTO_TCP_VERSION_MINOR 0
217 /* Acceptor connection request */
218 typedef struct lnet_acceptor_connreq {
219 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
220 __u32 acr_version; /* protocol version */
221 __u64 acr_nid; /* target NID */
222 } WIRE_ATTR lnet_acceptor_connreq_t;
224 #define LNET_PROTO_ACCEPTOR_VERSION 1
226 typedef struct lnet_counters {
238 } WIRE_ATTR lnet_counters_t;
240 #define LNET_NI_STATUS_UP 0x15aac0de
241 #define LNET_NI_STATUS_DOWN 0xdeadface
242 #define LNET_NI_STATUS_INVALID 0x00000000
244 struct lnet_ni_status {
251 * NB: value of these features equal to LNET_PROTO_PING_VERSION_x
252 * of old LNet, so there shouldn't be any compatibility issue
254 #define LNET_PING_FEAT_INVAL (0) /* no feature */
255 #define LNET_PING_FEAT_BASE (1 << 0) /* just a ping */
256 #define LNET_PING_FEAT_NI_STATUS (1 << 1) /* return NI status */
257 #define LNET_PING_FEAT_RTE_DISABLED (1 << 2) /* Routing enabled */
258 #define LNET_PING_FEAT_MULTI_RAIL (1 << 3) /* Multi-Rail aware */
259 #define LNET_PING_FEAT_DISCOVERY (1 << 4) /* Supports Discovery */
262 * All ping feature bits fit to hit the wire.
263 * In lnet_assert_wire_constants() this is compared against its open-coded
264 * value, and in lnet_ping_target_update() it is used to verify that no
265 * unknown bits have been set.
266 * New feature bits can be added, just be aware that this does change the
267 * over-the-wire protocol.
269 #define LNET_PING_FEAT_BITS (LNET_PING_FEAT_BASE | \
270 LNET_PING_FEAT_NI_STATUS | \
271 LNET_PING_FEAT_RTE_DISABLED | \
272 LNET_PING_FEAT_MULTI_RAIL | \
273 LNET_PING_FEAT_DISCOVERY)
275 struct lnet_ping_info {
280 struct lnet_ni_status pi_ni[0];
283 #define LNET_PING_INFO_SIZE(NNIDS) \
284 offsetof(struct lnet_ping_info, pi_ni[NNIDS])
285 #define LNET_PING_INFO_LONI(PINFO) ((PINFO)->pi_ni[0].ns_nid)
286 #define LNET_PING_INFO_SEQNO(PINFO) ((PINFO)->pi_ni[0].ns_status)
289 * This is a hard-coded limit on the number of interfaces supported by
290 * the interface bonding implemented by the ksocknal LND. It must be
291 * defined here because it is used in LNet data structures that are
292 * common to all LNDs.
294 #define LNET_INTERFACES_NUM 16
296 /* The minimum number of interfaces per node supported by LNet. */
297 #define LNET_INTERFACES_MIN 16
298 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
299 #define LNET_INTERFACES_MAX_DEFAULT 200
302 * Objects maintained by the LNet are accessed through handles. Handle types
303 * have names of the form lnet_handle_xx, where xx is one of the two letter
304 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
305 * 'me' for match entry). Each type of object is given a unique handle type
306 * to enhance type checking.
308 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
310 typedef struct lnet_handle_eq {
315 * Invalidate eq handle \a h.
317 static inline void LNetInvalidateEQHandle(struct lnet_handle_eq *h)
319 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
323 * Check whether eq handle \a h is invalid.
325 * \return 1 if handle is invalid, 0 if valid.
327 static inline int LNetEQHandleIsInvalid(struct lnet_handle_eq h)
329 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
332 typedef struct lnet_handle_md {
337 * Invalidate md handle \a h.
339 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
341 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
345 * Check whether eq handle \a h is invalid.
347 * \return 1 if handle is invalid, 0 if valid.
349 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
351 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
354 typedef struct lnet_handle_me {
361 typedef struct lnet_process_id {
369 /** \addtogroup lnet_me
373 * Specifies whether the match entry or memory descriptor should be unlinked
374 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
376 typedef enum lnet_unlink {
382 * Values of the type enum lnet_ins_pos are used to control where a new match
383 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
384 * entry before the current entry or before the head of the list. The value
385 * LNET_INS_AFTER is used to insert the new entry after the current entry
386 * or after the last item in the list.
388 typedef enum lnet_ins_pos {
389 /** insert ME before current position or head of the list */
391 /** insert ME after current position or tail of the list */
393 /** attach ME at tail of local CPU partition ME list */
399 /** \addtogroup lnet_md
403 * Defines the visible parts of a memory descriptor. Values of this type
404 * are used to initialize memory descriptors.
406 typedef struct lnet_md {
408 * Specify the memory region associated with the memory descriptor.
409 * If the options field has:
410 * - LNET_MD_KIOV bit set: The start field points to the starting
411 * address of an array of lnet_kiov_t and the length field specifies
412 * the number of entries in the array. The length can't be bigger
413 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
414 * fragments that are not necessarily mapped in virtal memory.
415 * - LNET_MD_IOVEC bit set: The start field points to the starting
416 * address of an array of struct kvec and the length field specifies
417 * the number of entries in the array. The length can't be bigger
418 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
419 * that have virtual addresses.
420 * - Otherwise: The memory region is contiguous. The start field
421 * specifies the starting address for the memory region and the
422 * length field specifies its length.
424 * When the memory region is fragmented, all fragments but the first
425 * one must start on page boundary, and all but the last must end on
431 * Specifies the maximum number of operations that can be performed
432 * on the memory descriptor. An operation is any action that could
433 * possibly generate an event. In the usual case, the threshold value
434 * is decremented for each operation on the MD. When the threshold
435 * drops to zero, the MD becomes inactive and does not respond to
436 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
437 * there is no bound on the number of operations that may be applied
442 * Specifies the largest incoming request that the memory descriptor
443 * should respond to. When the unused portion of a MD (length -
444 * local offset) falls below this value, the MD becomes inactive and
445 * does not respond to further operations. This value is only used
446 * if the LNET_MD_MAX_SIZE option is set.
450 * Specifies the behavior of the memory descriptor. A bitwise OR
451 * of the following values can be used:
452 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
453 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
454 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
455 * region is provided by the incoming request. By default, the
456 * offset is maintained locally. When maintained locally, the
457 * offset is incremented by the length of the request so that
458 * the next operation (PUT or GET) will access the next part of
459 * the memory region. Note that only one offset variable exists
460 * per memory descriptor. If both PUT and GET operations are
461 * performed on a memory descriptor, the offset is updated each time.
462 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
463 * be reduced to match the memory available in the region (determined
464 * by subtracting the offset from the length of the memory region).
465 * By default, if the length in the incoming operation is greater
466 * than the amount of memory available, the operation is rejected.
467 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
468 * incoming PUT operations, even if requested. By default,
469 * acknowledgments are sent for PUT operations that request an
470 * acknowledgment. Acknowledgments are never sent for GET operations.
471 * The data sent in the REPLY serves as an implicit acknowledgment.
472 * - LNET_MD_KIOV: The start and length fields specify an array of
474 * - LNET_MD_IOVEC: The start and length fields specify an array of
476 * - LNET_MD_MAX_SIZE: The max_size field is valid.
477 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
480 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
481 * capability for memory descriptors. They can't be both set.
482 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
483 * region (i.e. sum of all fragment lengths) must not be less than
486 unsigned int options;
488 * A user-specified value that is associated with the memory
489 * descriptor. The value does not need to be a pointer, but must fit
490 * in the space used by a pointer. This value is recorded in events
491 * associated with operations on this MD.
495 * A handle for the event queue used to log the operations performed on
496 * the memory region. If this argument is a NULL handle (i.e. nullified
497 * by LNetInvalidateHandle()), operations performed on this memory
498 * descriptor are not logged.
500 struct lnet_handle_eq eq_handle;
502 * The bulk MD handle which was registered to describe the buffers
503 * either to be used to transfer data to the peer or receive data
504 * from the peer. This allows LNet to properly determine the NUMA
505 * node on which the memory was allocated and use that to select the
506 * nearest local network interface. This value is only used
507 * if the LNET_MD_BULK_HANDLE option is set.
509 struct lnet_handle_md bulk_handle;
512 /* Max Transfer Unit (minimum supported everywhere).
513 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
514 * these limits are system wide and not interface-local. */
515 #define LNET_MTU_BITS 20
516 #define LNET_MTU (1 << LNET_MTU_BITS)
518 /** limit on the number of fragments in discontiguous MDs */
519 #define LNET_MAX_IOV 256
522 * Options for the MD structure. See struct lnet_md::options.
524 #define LNET_MD_OP_PUT (1 << 0)
525 /** See struct lnet_md::options. */
526 #define LNET_MD_OP_GET (1 << 1)
527 /** See struct lnet_md::options. */
528 #define LNET_MD_MANAGE_REMOTE (1 << 2)
529 /* unused (1 << 3) */
530 /** See struct lnet_md::options. */
531 #define LNET_MD_TRUNCATE (1 << 4)
532 /** See struct lnet_md::options. */
533 #define LNET_MD_ACK_DISABLE (1 << 5)
534 /** See struct lnet_md::options. */
535 #define LNET_MD_IOVEC (1 << 6)
536 /** See struct lnet_md::options. */
537 #define LNET_MD_MAX_SIZE (1 << 7)
538 /** See struct lnet_md::options. */
539 #define LNET_MD_KIOV (1 << 8)
540 /** See struct lnet_md::options. */
541 #define LNET_MD_BULK_HANDLE (1 << 9)
543 /* For compatibility with Cray Portals */
544 #define LNET_MD_PHYS 0
546 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
547 #define LNET_MD_THRESH_INF (-1)
550 * A page-based fragment of a MD.
553 /** Pointer to the page where the fragment resides */
554 struct page *kiov_page;
555 /** Length in bytes of the fragment */
556 unsigned int kiov_len;
558 * Starting offset of the fragment within the page. Note that the
559 * end of the fragment must not pass the end of the page; i.e.,
560 * kiov_len + kiov_offset <= PAGE_SIZE.
562 unsigned int kiov_offset;
566 /** \addtogroup lnet_eq
570 * Six types of events can be logged in an event queue.
572 typedef enum lnet_event_kind {
573 /** An incoming GET operation has completed on the MD. */
576 * An incoming PUT operation has completed on the MD. The
577 * underlying layers will not alter the memory (on behalf of this
578 * operation) once this event has been logged.
582 * A REPLY operation has completed. This event is logged after the
583 * data (if any) from the REPLY has been written into the MD.
586 /** An acknowledgment has been received. */
589 * An outgoing send (PUT or GET) operation has completed. This event
590 * is logged after the entire buffer has been sent and it is safe for
591 * the caller to reuse the buffer.
594 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
595 * happen even when the message has not yet been put out on wire.
596 * - It's unsafe to assume that in an outgoing GET operation
597 * the LNET_EVENT_SEND event would happen before the
598 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
599 * LNET_EVENT_ACK events in an outgoing PUT operation.
603 * A MD has been unlinked. Note that LNetMDUnlink() does not
604 * necessarily trigger an LNET_EVENT_UNLINK event.
610 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
613 * Information about an event on a MD.
615 typedef struct lnet_event {
616 /** The identifier (nid, pid) of the target. */
617 struct lnet_process_id target;
618 /** The identifier (nid, pid) of the initiator. */
619 struct lnet_process_id initiator;
620 /** The source NID on the initiator. */
621 struct lnet_process_id source;
623 * The NID of the immediate sender. If the request has been forwarded
624 * by routers, this is the NID of the last hop; otherwise it's the
625 * same as the source.
628 /** Indicates the type of the event. */
629 enum lnet_event_kind type;
630 /** The portal table index specified in the request */
631 unsigned int pt_index;
632 /** A copy of the match bits specified in the request. */
634 /** The length (in bytes) specified in the request. */
635 unsigned int rlength;
637 * The length (in bytes) of the data that was manipulated by the
638 * operation. For truncated operations, the manipulated length will be
639 * the number of bytes specified by the MD (possibly with an offset,
640 * see struct lnet_md). For all other operations, the manipulated length
641 * will be the length of the requested operation, i.e. rlength.
643 unsigned int mlength;
645 * The handle to the MD associated with the event. The handle may be
646 * invalid if the MD has been unlinked.
648 struct lnet_handle_md md_handle;
650 * A snapshot of the state of the MD immediately after the event has
651 * been processed. In particular, the threshold field in md will
652 * reflect the value of the threshold after the operation occurred.
656 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
661 * Indicates the completion status of the operation. It's 0 for
662 * successful operations, otherwise it's an error code.
666 * Indicates whether the MD has been unlinked. Note that:
667 * - An event with unlinked set is the last event on the MD.
668 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
673 * The displacement (in bytes) into the memory region that the
674 * operation used. The offset can be determined by the operation for
675 * a remote managed MD or by the local MD.
676 * \see struct lnet_md::options
680 * The sequence number for this event. Sequence numbers are unique
683 volatile unsigned long sequence;
687 * Event queue handler function type.
689 * The EQ handler runs for each event that is deposited into the EQ. The
690 * handler is supplied with a pointer to the event that triggered the
691 * handler invocation.
693 * The handler must not block, must be reentrant, and must not call any LNet
694 * API functions. It should return as quickly as possible.
696 typedef void (*lnet_eq_handler_t)(struct lnet_event *event);
697 #define LNET_EQ_HANDLER_NONE NULL
700 /** \addtogroup lnet_data
704 * Specify whether an acknowledgment should be sent by target when the PUT
705 * operation completes (i.e., when the data has been written to a MD of the
708 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
709 * by which acknowledgments can be disabled for a MD.
711 typedef enum lnet_ack_req {
712 /** Request an acknowledgment */
714 /** Request that no acknowledgment should be generated. */