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 __LNET_TYPES_H__
34 #define __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;
62 * ID of a process in a node. Shortened as PID to distinguish from
63 * lnet_process_id_t, the global process ID.
65 typedef __u32 lnet_pid_t;
67 /** wildcard NID that matches any end-point address */
68 #define LNET_NID_ANY ((lnet_nid_t) -1)
69 /** wildcard PID that matches any lnet_pid_t */
70 #define LNET_PID_ANY ((lnet_pid_t) -1)
72 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
73 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
74 #define LNET_PID_LUSTRE 12345
76 /* how an LNET NID encodes net:address */
77 /** extract the address part of an lnet_nid_t */
79 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
81 return nid & 0xffffffff;
84 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
86 return (nid >> 32) & 0xffffffff;
89 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
91 return (((__u64)net) << 32) | addr;
94 static inline __u32 LNET_NETNUM(__u32 net)
99 static inline __u32 LNET_NETTYP(__u32 net)
101 return (net >> 16) & 0xffff;
104 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
106 return (type << 16) | num;
109 #define WIRE_ATTR __attribute__((packed))
111 /* Packed version of lnet_process_id_t to transfer via network */
112 typedef struct lnet_process_id_packed {
114 lnet_pid_t pid; /* node id / process id */
115 } WIRE_ATTR lnet_process_id_packed_t;
117 /* The wire handle's interface cookie only matches one network interface in
118 * one epoch (i.e. new cookie when the interface restarts or the node
119 * reboots). The object cookie only matches one object on that interface
120 * during that object's lifetime (i.e. no cookie re-use). */
121 typedef struct lnet_handle_wire {
122 __u64 wh_interface_cookie;
123 __u64 wh_object_cookie;
124 } WIRE_ATTR lnet_handle_wire_t;
134 /* The variant fields of the portals message header are aligned on an 8
135 * byte boundary in the message header. Note that all types used in these
136 * wire structs MUST be fixed size and the smaller types are placed at the
139 struct lnet_handle_wire dst_wmd;
145 struct lnet_handle_wire ack_wmd;
153 struct lnet_handle_wire return_wmd;
161 struct lnet_handle_wire dst_wmd;
169 typedef struct lnet_hdr {
174 __u32 type; /* lnet_msg_type_t */
175 __u32 payload_length; /* payload data to follow */
176 /*<------__u64 aligned------->*/
181 struct lnet_reply reply;
182 struct lnet_hello hello;
184 } WIRE_ATTR lnet_hdr_t;
186 /* A HELLO message contains a magic number and protocol version
187 * code in the header's dest_nid, the peer's NID in the src_nid, and
188 * LNET_MSG_HELLO in the type field. All other common fields are zero
189 * (including payload_size; i.e. no payload).
190 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
191 * running the same protocol and to find out its NID. These LNDs should
192 * exchange HELLO messages when a connection is first established. Individual
193 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
195 typedef struct lnet_magicversion {
196 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
197 __u16 version_major; /* increment on incompatible change */
198 __u16 version_minor; /* increment on compatible change */
199 } WIRE_ATTR lnet_magicversion_t;
201 /* PROTO MAGIC for LNDs */
202 #define LNET_PROTO_IB_MAGIC 0x0be91b91
203 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
204 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
205 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
206 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
208 /* Placeholder for a future "unified" protocol across all LNDs */
209 /* Current LNDs that receive a request with this magic will respond
210 * with a "stub" reply using their current protocol */
211 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
213 #define LNET_PROTO_TCP_VERSION_MAJOR 1
214 #define LNET_PROTO_TCP_VERSION_MINOR 0
216 /* Acceptor connection request */
217 typedef struct lnet_acceptor_connreq {
218 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
219 __u32 acr_version; /* protocol version */
220 __u64 acr_nid; /* target NID */
221 } WIRE_ATTR lnet_acceptor_connreq_t;
223 #define LNET_PROTO_ACCEPTOR_VERSION 1
225 typedef struct lnet_counters {
237 } WIRE_ATTR lnet_counters_t;
239 #define LNET_NI_STATUS_UP 0x15aac0de
240 #define LNET_NI_STATUS_DOWN 0xdeadface
241 #define LNET_NI_STATUS_INVALID 0x00000000
243 #define LNET_MAX_INTERFACES 16
246 * Objects maintained by the LNet are accessed through handles. Handle types
247 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
248 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
249 * 'me' for match entry).
250 * Each type of object is given a unique handle type to enhance type checking.
251 * The type lnet_handle_any_t can be used when a generic handle is needed.
252 * Every handle value can be converted into a value of type lnet_handle_any_t
253 * without loss of information.
259 typedef lnet_handle_any_t lnet_handle_eq_t;
260 typedef lnet_handle_any_t lnet_handle_md_t;
261 typedef lnet_handle_any_t lnet_handle_me_t;
263 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
266 * Invalidate handle \a h.
268 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
270 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
274 * Compare handles \a h1 and \a h2.
276 * \return 1 if handles are equal, 0 if otherwise.
278 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
280 return (h1.cookie == h2.cookie);
284 * Check whether handle \a h is invalid.
286 * \return 1 if handle is invalid, 0 if valid.
288 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
290 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
304 /** \addtogroup lnet_me
308 * Specifies whether the match entry or memory descriptor should be unlinked
309 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
317 * Values of the type lnet_ins_pos_t are used to control where a new match
318 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
319 * entry before the current entry or before the head of the list. The value
320 * LNET_INS_AFTER is used to insert the new entry after the current entry
321 * or after the last item in the list.
324 /** insert ME before current position or head of the list */
326 /** insert ME after current position or tail of the list */
328 /** attach ME at tail of local CPU partition ME list */
334 /** \addtogroup lnet_md
338 * Defines the visible parts of a memory descriptor. Values of this type
339 * are used to initialize memory descriptors.
343 * Specify the memory region associated with the memory descriptor.
344 * If the options field has:
345 * - LNET_MD_KIOV bit set: The start field points to the starting
346 * address of an array of lnet_kiov_t and the length field specifies
347 * the number of entries in the array. The length can't be bigger
348 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
349 * fragments that are not necessarily mapped in virtal memory.
350 * - LNET_MD_IOVEC bit set: The start field points to the starting
351 * address of an array of struct kvec and the length field specifies
352 * the number of entries in the array. The length can't be bigger
353 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
354 * that have virtual addresses.
355 * - Otherwise: The memory region is contiguous. The start field
356 * specifies the starting address for the memory region and the
357 * length field specifies its length.
359 * When the memory region is fragmented, all fragments but the first
360 * one must start on page boundary, and all but the last must end on
366 * Specifies the maximum number of operations that can be performed
367 * on the memory descriptor. An operation is any action that could
368 * possibly generate an event. In the usual case, the threshold value
369 * is decremented for each operation on the MD. When the threshold
370 * drops to zero, the MD becomes inactive and does not respond to
371 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
372 * there is no bound on the number of operations that may be applied
377 * Specifies the largest incoming request that the memory descriptor
378 * should respond to. When the unused portion of a MD (length -
379 * local offset) falls below this value, the MD becomes inactive and
380 * does not respond to further operations. This value is only used
381 * if the LNET_MD_MAX_SIZE option is set.
385 * Specifies the behavior of the memory descriptor. A bitwise OR
386 * of the following values can be used:
387 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
388 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
389 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
390 * region is provided by the incoming request. By default, the
391 * offset is maintained locally. When maintained locally, the
392 * offset is incremented by the length of the request so that
393 * the next operation (PUT or GET) will access the next part of
394 * the memory region. Note that only one offset variable exists
395 * per memory descriptor. If both PUT and GET operations are
396 * performed on a memory descriptor, the offset is updated each time.
397 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
398 * be reduced to match the memory available in the region (determined
399 * by subtracting the offset from the length of the memory region).
400 * By default, if the length in the incoming operation is greater
401 * than the amount of memory available, the operation is rejected.
402 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
403 * incoming PUT operations, even if requested. By default,
404 * acknowledgments are sent for PUT operations that request an
405 * acknowledgment. Acknowledgments are never sent for GET operations.
406 * The data sent in the REPLY serves as an implicit acknowledgment.
407 * - LNET_MD_KIOV: The start and length fields specify an array of
409 * - LNET_MD_IOVEC: The start and length fields specify an array of
411 * - LNET_MD_MAX_SIZE: The max_size field is valid.
414 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
415 * capability for memory descriptors. They can't be both set.
416 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
417 * region (i.e. sum of all fragment lengths) must not be less than
420 unsigned int options;
422 * A user-specified value that is associated with the memory
423 * descriptor. The value does not need to be a pointer, but must fit
424 * in the space used by a pointer. This value is recorded in events
425 * associated with operations on this MD.
429 * A handle for the event queue used to log the operations performed on
430 * the memory region. If this argument is a NULL handle (i.e. nullified
431 * by LNetInvalidateHandle()), operations performed on this memory
432 * descriptor are not logged.
434 lnet_handle_eq_t eq_handle;
437 /* Max Transfer Unit (minimum supported everywhere).
438 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
439 * these limits are system wide and not interface-local. */
440 #define LNET_MTU_BITS 20
441 #define LNET_MTU (1 << LNET_MTU_BITS)
443 /** limit on the number of fragments in discontiguous MDs */
444 #define LNET_MAX_IOV 256
447 * Options for the MD structure. See lnet_md_t::options.
449 #define LNET_MD_OP_PUT (1 << 0)
450 /** See lnet_md_t::options. */
451 #define LNET_MD_OP_GET (1 << 1)
452 /** See lnet_md_t::options. */
453 #define LNET_MD_MANAGE_REMOTE (1 << 2)
454 /* unused (1 << 3) */
455 /** See lnet_md_t::options. */
456 #define LNET_MD_TRUNCATE (1 << 4)
457 /** See lnet_md_t::options. */
458 #define LNET_MD_ACK_DISABLE (1 << 5)
459 /** See lnet_md_t::options. */
460 #define LNET_MD_IOVEC (1 << 6)
461 /** See lnet_md_t::options. */
462 #define LNET_MD_MAX_SIZE (1 << 7)
463 /** See lnet_md_t::options. */
464 #define LNET_MD_KIOV (1 << 8)
466 /* For compatibility with Cray Portals */
467 #define LNET_MD_PHYS 0
469 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
470 #define LNET_MD_THRESH_INF (-1)
472 /* NB lustre portals uses struct iovec internally! */
473 typedef struct iovec lnet_md_iovec_t;
476 * A page-based fragment of a MD.
479 /** Pointer to the page where the fragment resides */
480 struct page *kiov_page;
481 /** Length in bytes of the fragment */
482 unsigned int kiov_len;
484 * Starting offset of the fragment within the page. Note that the
485 * end of the fragment must not pass the end of the page; i.e.,
486 * kiov_len + kiov_offset <= PAGE_SIZE.
488 unsigned int kiov_offset;
492 /** \addtogroup lnet_eq
496 * Six types of events can be logged in an event queue.
499 /** An incoming GET operation has completed on the MD. */
502 * An incoming PUT operation has completed on the MD. The
503 * underlying layers will not alter the memory (on behalf of this
504 * operation) once this event has been logged.
508 * A REPLY operation has completed. This event is logged after the
509 * data (if any) from the REPLY has been written into the MD.
512 /** An acknowledgment has been received. */
515 * An outgoing send (PUT or GET) operation has completed. This event
516 * is logged after the entire buffer has been sent and it is safe for
517 * the caller to reuse the buffer.
520 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
521 * happen even when the message has not yet been put out on wire.
522 * - It's unsafe to assume that in an outgoing GET operation
523 * the LNET_EVENT_SEND event would happen before the
524 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
525 * LNET_EVENT_ACK events in an outgoing PUT operation.
529 * A MD has been unlinked. Note that LNetMDUnlink() does not
530 * necessarily trigger an LNET_EVENT_UNLINK event.
536 #define LNET_SEQ_BASETYPE long
537 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
538 #define LNET_SEQ_GT(a, b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
541 * Information about an event on a MD.
544 /** The identifier (nid, pid) of the target. */
545 lnet_process_id_t target;
546 /** The identifier (nid, pid) of the initiator. */
547 lnet_process_id_t initiator;
548 /** The source NID on the initiator. */
549 lnet_process_id_t source;
551 * The NID of the immediate sender. If the request has been forwarded
552 * by routers, this is the NID of the last hop; otherwise it's the
553 * same as the source.
556 /** Indicates the type of the event. */
557 lnet_event_kind_t type;
558 /** The portal table index specified in the request */
559 unsigned int pt_index;
560 /** A copy of the match bits specified in the request. */
562 /** The length (in bytes) specified in the request. */
563 unsigned int rlength;
565 * The length (in bytes) of the data that was manipulated by the
566 * operation. For truncated operations, the manipulated length will be
567 * the number of bytes specified by the MD (possibly with an offset,
568 * see lnet_md_t). For all other operations, the manipulated length
569 * will be the length of the requested operation, i.e. rlength.
571 unsigned int mlength;
573 * The handle to the MD associated with the event. The handle may be
574 * invalid if the MD has been unlinked.
576 lnet_handle_md_t md_handle;
578 * A snapshot of the state of the MD immediately after the event has
579 * been processed. In particular, the threshold field in md will
580 * reflect the value of the threshold after the operation occurred.
584 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
589 * Indicates the completion status of the operation. It's 0 for
590 * successful operations, otherwise it's an error code.
594 * Indicates whether the MD has been unlinked. Note that:
595 * - An event with unlinked set is the last event on the MD.
596 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
601 * The displacement (in bytes) into the memory region that the
602 * operation used. The offset can be determined by the operation for
603 * a remote managed MD or by the local MD.
604 * \see lnet_md_t::options
608 * The sequence number for this event. Sequence numbers are unique
611 volatile lnet_seq_t sequence;
615 * Event queue handler function type.
617 * The EQ handler runs for each event that is deposited into the EQ. The
618 * handler is supplied with a pointer to the event that triggered the
619 * handler invocation.
621 * The handler must not block, must be reentrant, and must not call any LNet
622 * API functions. It should return as quickly as possible.
624 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
625 #define LNET_EQ_HANDLER_NONE NULL
628 /** \addtogroup lnet_data
632 * Specify whether an acknowledgment should be sent by target when the PUT
633 * operation completes (i.e., when the data has been written to a MD of the
636 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
637 * acknowledgments can be disabled for a MD.
640 /** Request an acknowledgment */
642 /** Request that no acknowledgment should be generated. */