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
245 * This is a hard-coded limit on the number of interfaces supported by
246 * the interface bonding implemented by the ksocknal LND. It must be
247 * defined here because it is used in LNet data structures that are
248 * common to all LNDs.
250 #define LNET_INTERFACES_NUM 16
252 /* The minimum number of interfaces per node supported by LNet. */
253 #define LNET_INTERFACES_MIN 16
254 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
255 #define LNET_INTERFACES_MAX_DEFAULT 200
258 * Objects maintained by the LNet are accessed through handles. Handle types
259 * have names of the form lnet_handle_xx, where xx is one of the two letter
260 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
261 * 'me' for match entry). Each type of object is given a unique handle type
262 * to enhance type checking.
264 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
266 typedef struct lnet_handle_eq {
271 * Invalidate eq handle \a h.
273 static inline void LNetInvalidateEQHandle(struct lnet_handle_eq *h)
275 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
279 * Check whether eq handle \a h is invalid.
281 * \return 1 if handle is invalid, 0 if valid.
283 static inline int LNetEQHandleIsInvalid(struct lnet_handle_eq h)
285 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
288 typedef struct lnet_handle_md {
293 * Invalidate md handle \a h.
295 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
297 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
301 * Check whether eq handle \a h is invalid.
303 * \return 1 if handle is invalid, 0 if valid.
305 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
307 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
310 typedef struct lnet_handle_me {
317 typedef struct lnet_process_id {
325 /** \addtogroup lnet_me
329 * Specifies whether the match entry or memory descriptor should be unlinked
330 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
332 typedef enum lnet_unlink {
338 * Values of the type enum lnet_ins_pos are used to control where a new match
339 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
340 * entry before the current entry or before the head of the list. The value
341 * LNET_INS_AFTER is used to insert the new entry after the current entry
342 * or after the last item in the list.
344 typedef enum lnet_ins_pos {
345 /** insert ME before current position or head of the list */
347 /** insert ME after current position or tail of the list */
349 /** attach ME at tail of local CPU partition ME list */
355 /** \addtogroup lnet_md
359 * Defines the visible parts of a memory descriptor. Values of this type
360 * are used to initialize memory descriptors.
362 typedef struct lnet_md {
364 * Specify the memory region associated with the memory descriptor.
365 * If the options field has:
366 * - LNET_MD_KIOV bit set: The start field points to the starting
367 * address of an array of lnet_kiov_t and the length field specifies
368 * the number of entries in the array. The length can't be bigger
369 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
370 * fragments that are not necessarily mapped in virtal memory.
371 * - LNET_MD_IOVEC bit set: The start field points to the starting
372 * address of an array of struct kvec and the length field specifies
373 * the number of entries in the array. The length can't be bigger
374 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
375 * that have virtual addresses.
376 * - Otherwise: The memory region is contiguous. The start field
377 * specifies the starting address for the memory region and the
378 * length field specifies its length.
380 * When the memory region is fragmented, all fragments but the first
381 * one must start on page boundary, and all but the last must end on
387 * Specifies the maximum number of operations that can be performed
388 * on the memory descriptor. An operation is any action that could
389 * possibly generate an event. In the usual case, the threshold value
390 * is decremented for each operation on the MD. When the threshold
391 * drops to zero, the MD becomes inactive and does not respond to
392 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
393 * there is no bound on the number of operations that may be applied
398 * Specifies the largest incoming request that the memory descriptor
399 * should respond to. When the unused portion of a MD (length -
400 * local offset) falls below this value, the MD becomes inactive and
401 * does not respond to further operations. This value is only used
402 * if the LNET_MD_MAX_SIZE option is set.
406 * Specifies the behavior of the memory descriptor. A bitwise OR
407 * of the following values can be used:
408 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
409 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
410 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
411 * region is provided by the incoming request. By default, the
412 * offset is maintained locally. When maintained locally, the
413 * offset is incremented by the length of the request so that
414 * the next operation (PUT or GET) will access the next part of
415 * the memory region. Note that only one offset variable exists
416 * per memory descriptor. If both PUT and GET operations are
417 * performed on a memory descriptor, the offset is updated each time.
418 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
419 * be reduced to match the memory available in the region (determined
420 * by subtracting the offset from the length of the memory region).
421 * By default, if the length in the incoming operation is greater
422 * than the amount of memory available, the operation is rejected.
423 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
424 * incoming PUT operations, even if requested. By default,
425 * acknowledgments are sent for PUT operations that request an
426 * acknowledgment. Acknowledgments are never sent for GET operations.
427 * The data sent in the REPLY serves as an implicit acknowledgment.
428 * - LNET_MD_KIOV: The start and length fields specify an array of
430 * - LNET_MD_IOVEC: The start and length fields specify an array of
432 * - LNET_MD_MAX_SIZE: The max_size field is valid.
433 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
436 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
437 * capability for memory descriptors. They can't be both set.
438 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
439 * region (i.e. sum of all fragment lengths) must not be less than
442 unsigned int options;
444 * A user-specified value that is associated with the memory
445 * descriptor. The value does not need to be a pointer, but must fit
446 * in the space used by a pointer. This value is recorded in events
447 * associated with operations on this MD.
451 * A handle for the event queue used to log the operations performed on
452 * the memory region. If this argument is a NULL handle (i.e. nullified
453 * by LNetInvalidateHandle()), operations performed on this memory
454 * descriptor are not logged.
456 struct lnet_handle_eq eq_handle;
458 * The bulk MD handle which was registered to describe the buffers
459 * either to be used to transfer data to the peer or receive data
460 * from the peer. This allows LNet to properly determine the NUMA
461 * node on which the memory was allocated and use that to select the
462 * nearest local network interface. This value is only used
463 * if the LNET_MD_BULK_HANDLE option is set.
465 struct lnet_handle_md bulk_handle;
468 /* Max Transfer Unit (minimum supported everywhere).
469 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
470 * 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 struct lnet_md::options.
480 #define LNET_MD_OP_PUT (1 << 0)
481 /** See struct lnet_md::options. */
482 #define LNET_MD_OP_GET (1 << 1)
483 /** See struct lnet_md::options. */
484 #define LNET_MD_MANAGE_REMOTE (1 << 2)
485 /* unused (1 << 3) */
486 /** See struct lnet_md::options. */
487 #define LNET_MD_TRUNCATE (1 << 4)
488 /** See struct lnet_md::options. */
489 #define LNET_MD_ACK_DISABLE (1 << 5)
490 /** See struct lnet_md::options. */
491 #define LNET_MD_IOVEC (1 << 6)
492 /** See struct lnet_md::options. */
493 #define LNET_MD_MAX_SIZE (1 << 7)
494 /** See struct lnet_md::options. */
495 #define LNET_MD_KIOV (1 << 8)
496 /** See struct lnet_md::options. */
497 #define LNET_MD_BULK_HANDLE (1 << 9)
499 /* For compatibility with Cray Portals */
500 #define LNET_MD_PHYS 0
502 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
503 #define LNET_MD_THRESH_INF (-1)
506 * A page-based fragment of a MD.
509 /** Pointer to the page where the fragment resides */
510 struct page *kiov_page;
511 /** Length in bytes of the fragment */
512 unsigned int kiov_len;
514 * Starting offset of the fragment within the page. Note that the
515 * end of the fragment must not pass the end of the page; i.e.,
516 * kiov_len + kiov_offset <= PAGE_SIZE.
518 unsigned int kiov_offset;
522 /** \addtogroup lnet_eq
526 * Six types of events can be logged in an event queue.
528 typedef enum lnet_event_kind {
529 /** An incoming GET operation has completed on the MD. */
532 * An incoming PUT operation has completed on the MD. The
533 * underlying layers will not alter the memory (on behalf of this
534 * operation) once this event has been logged.
538 * A REPLY operation has completed. This event is logged after the
539 * data (if any) from the REPLY has been written into the MD.
542 /** An acknowledgment has been received. */
545 * An outgoing send (PUT or GET) operation has completed. This event
546 * is logged after the entire buffer has been sent and it is safe for
547 * the caller to reuse the buffer.
550 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
551 * happen even when the message has not yet been put out on wire.
552 * - It's unsafe to assume that in an outgoing GET operation
553 * the LNET_EVENT_SEND event would happen before the
554 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
555 * LNET_EVENT_ACK events in an outgoing PUT operation.
559 * A MD has been unlinked. Note that LNetMDUnlink() does not
560 * necessarily trigger an LNET_EVENT_UNLINK event.
566 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
569 * Information about an event on a MD.
571 typedef struct lnet_event {
572 /** The identifier (nid, pid) of the target. */
573 struct lnet_process_id target;
574 /** The identifier (nid, pid) of the initiator. */
575 struct lnet_process_id initiator;
576 /** The source NID on the initiator. */
577 struct lnet_process_id source;
579 * The NID of the immediate sender. If the request has been forwarded
580 * by routers, this is the NID of the last hop; otherwise it's the
581 * same as the source.
584 /** Indicates the type of the event. */
585 enum lnet_event_kind type;
586 /** The portal table index specified in the request */
587 unsigned int pt_index;
588 /** A copy of the match bits specified in the request. */
590 /** The length (in bytes) specified in the request. */
591 unsigned int rlength;
593 * The length (in bytes) of the data that was manipulated by the
594 * operation. For truncated operations, the manipulated length will be
595 * the number of bytes specified by the MD (possibly with an offset,
596 * see struct lnet_md). For all other operations, the manipulated length
597 * will be the length of the requested operation, i.e. rlength.
599 unsigned int mlength;
601 * The handle to the MD associated with the event. The handle may be
602 * invalid if the MD has been unlinked.
604 struct lnet_handle_md md_handle;
606 * A snapshot of the state of the MD immediately after the event has
607 * been processed. In particular, the threshold field in md will
608 * reflect the value of the threshold after the operation occurred.
612 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
617 * Indicates the completion status of the operation. It's 0 for
618 * successful operations, otherwise it's an error code.
622 * Indicates whether the MD has been unlinked. Note that:
623 * - An event with unlinked set is the last event on the MD.
624 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
629 * The displacement (in bytes) into the memory region that the
630 * operation used. The offset can be determined by the operation for
631 * a remote managed MD or by the local MD.
632 * \see struct lnet_md::options
636 * The sequence number for this event. Sequence numbers are unique
639 volatile unsigned long sequence;
643 * Event queue handler function type.
645 * The EQ handler runs for each event that is deposited into the EQ. The
646 * handler is supplied with a pointer to the event that triggered the
647 * handler invocation.
649 * The handler must not block, must be reentrant, and must not call any LNet
650 * API functions. It should return as quickly as possible.
652 typedef void (*lnet_eq_handler_t)(struct lnet_event *event);
653 #define LNET_EQ_HANDLER_NONE NULL
656 /** \addtogroup lnet_data
660 * Specify whether an acknowledgment should be sent by target when the PUT
661 * operation completes (i.e., when the data has been written to a MD of the
664 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
665 * by which acknowledgments can be disabled for a MD.
667 typedef enum lnet_ack_req {
668 /** Request an acknowledgment */
670 /** Request that no acknowledgment should be generated. */