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 <libcfs/types.h>
45 /** \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 __attribute__((packed))
116 /* Packed version of lnet_process_id_t to transfer via network */
119 lnet_pid_t pid; /* node id / process id */
120 } WIRE_ATTR lnet_process_id_packed_t;
122 /* The wire handle's interface cookie only matches one network interface in
123 * one epoch (i.e. new cookie when the interface restarts or the node
124 * reboots). The object cookie only matches one object on that interface
125 * during that object's lifetime (i.e. no cookie re-use). */
127 __u64 wh_interface_cookie;
128 __u64 wh_object_cookie;
129 } WIRE_ATTR lnet_handle_wire_t;
139 /* The variant fields of the portals message header are aligned on an 8
140 * byte boundary in the message header. Note that all types used in these
141 * wire structs MUST be fixed size and the smaller types are placed at the
143 typedef struct lnet_ack {
144 lnet_handle_wire_t dst_wmd;
147 } WIRE_ATTR lnet_ack_t;
149 typedef struct lnet_put {
150 lnet_handle_wire_t ack_wmd;
155 } WIRE_ATTR lnet_put_t;
157 typedef struct lnet_get {
158 lnet_handle_wire_t return_wmd;
163 } WIRE_ATTR lnet_get_t;
165 typedef struct lnet_reply {
166 lnet_handle_wire_t dst_wmd;
167 } WIRE_ATTR lnet_reply_t;
169 typedef struct lnet_hello {
172 } WIRE_ATTR lnet_hello_t;
179 __u32 type; /* lnet_msg_type_t */
180 __u32 payload_length; /* payload data to follow */
181 /*<------__u64 aligned------->*/
189 } WIRE_ATTR lnet_hdr_t;
191 /* A HELLO message contains a magic number and protocol version
192 * code in the header's dest_nid, the peer's NID in the src_nid, and
193 * LNET_MSG_HELLO in the type field. All other common fields are zero
194 * (including payload_size; i.e. no payload).
195 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
196 * running the same protocol and to find out its NID. These LNDs should
197 * exchange HELLO messages when a connection is first established. Individual
198 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
201 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
202 __u16 version_major; /* increment on incompatible change */
203 __u16 version_minor; /* increment on compatible change */
204 } WIRE_ATTR lnet_magicversion_t;
206 /* PROTO MAGIC for LNDs */
207 #define LNET_PROTO_IB_MAGIC 0x0be91b91
208 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
209 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
210 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
211 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
213 /* Placeholder for a future "unified" protocol across all LNDs */
214 /* Current LNDs that receive a request with this magic will respond with a
215 * * * "stub" reply using their current protocol */
216 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
218 #define LNET_PROTO_TCP_VERSION_MAJOR 1
219 #define LNET_PROTO_TCP_VERSION_MINOR 0
221 /* Acceptor connection request */
223 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
224 __u32 acr_version; /* protocol version */
225 __u64 acr_nid; /* target NID */
226 } WIRE_ATTR lnet_acceptor_connreq_t;
228 #define LNET_PROTO_ACCEPTOR_VERSION 1
230 typedef struct lnet_counters {
242 } WIRE_ATTR lnet_counters_t;
244 #define LNET_NI_STATUS_UP 0x15aac0de
245 #define LNET_NI_STATUS_DOWN 0xdeadface
246 #define LNET_NI_STATUS_INVALID 0x00000000
248 #define LNET_MAX_INTERFACES 16
251 * Objects maintained by the LNet are accessed through handles. Handle types
252 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
253 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
254 * 'me' for match entry).
255 * Each type of object is given a unique handle type to enhance type checking.
256 * The type lnet_handle_any_t can be used when a generic handle is needed.
257 * Every handle value can be converted into a value of type lnet_handle_any_t
258 * without loss of information.
264 typedef lnet_handle_any_t lnet_handle_eq_t;
265 typedef lnet_handle_any_t lnet_handle_md_t;
266 typedef lnet_handle_any_t lnet_handle_me_t;
268 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
271 * Invalidate handle \a h.
273 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
275 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
279 * Compare handles \a h1 and \a h2.
281 * \return 1 if handles are equal, 0 if otherwise.
283 static inline int LNetHandleIsEqual (lnet_handle_any_t h1, lnet_handle_any_t h2)
285 return (h1.cookie == h2.cookie);
289 * Check whether handle \a h is invalid.
291 * \return 1 if handle is invalid, 0 if valid.
293 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
295 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
309 /** \addtogroup lnet_me
313 * Specifies whether the match entry or memory descriptor should be unlinked
314 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
322 * Values of the type lnet_ins_pos_t are used to control where a new match
323 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
324 * entry before the current entry or before the head of the list. The value
325 * LNET_INS_AFTER is used to insert the new entry after the current entry
326 * or after the last item in the list.
329 /** insert ME before current position or head of the list */
331 /** insert ME after current position or tail of the list */
333 /** attach ME at tail of local CPU partition ME list */
339 /** \addtogroup lnet_md
343 * Defines the visible parts of a memory descriptor. Values of this type
344 * are used to initialize memory descriptors.
348 * Specify the memory region associated with the memory descriptor.
349 * If the options field has:
350 * - LNET_MD_KIOV bit set: The start field points to the starting
351 * address of an array of lnet_kiov_t 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 lnet_kiov_t is used to describe page-based
354 * fragments that are not necessarily mapped in virtal memory.
355 * - LNET_MD_IOVEC bit set: The start field points to the starting
356 * address of an array of struct iovec and the length field specifies
357 * the number of entries in the array. The length can't be bigger
358 * than LNET_MAX_IOV. The struct iovec is used to describe fragments
359 * that have virtual addresses.
360 * - Otherwise: The memory region is contiguous. The start field
361 * specifies the starting address for the memory region and the
362 * length field specifies its length.
364 * When the memory region is fragmented, all fragments but the first
365 * one must start on page boundary, and all but the last must end on
371 * Specifies the maximum number of operations that can be performed
372 * on the memory descriptor. An operation is any action that could
373 * possibly generate an event. In the usual case, the threshold value
374 * is decremented for each operation on the MD. When the threshold
375 * drops to zero, the MD becomes inactive and does not respond to
376 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
377 * there is no bound on the number of operations that may be applied
382 * Specifies the largest incoming request that the memory descriptor
383 * should respond to. When the unused portion of a MD (length -
384 * local offset) falls below this value, the MD becomes inactive and
385 * does not respond to further operations. This value is only used
386 * if the LNET_MD_MAX_SIZE option is set.
390 * Specifies the behavior of the memory descriptor. A bitwise OR
391 * of the following values can be used:
392 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
393 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
394 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
395 * region is provided by the incoming request. By default, the
396 * offset is maintained locally. When maintained locally, the
397 * offset is incremented by the length of the request so that
398 * the next operation (PUT or GET) will access the next part of
399 * the memory region. Note that only one offset variable exists
400 * per memory descriptor. If both PUT and GET operations are
401 * performed on a memory descriptor, the offset is updated each time.
402 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
403 * be reduced to match the memory available in the region (determined
404 * by subtracting the offset from the length of the memory region).
405 * By default, if the length in the incoming operation is greater
406 * than the amount of memory available, the operation is rejected.
407 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
408 * incoming PUT operations, even if requested. By default,
409 * acknowledgments are sent for PUT operations that request an
410 * acknowledgment. Acknowledgments are never sent for GET operations.
411 * The data sent in the REPLY serves as an implicit acknowledgment.
412 * - LNET_MD_KIOV: The start and length fields specify an array of
414 * - LNET_MD_IOVEC: The start and length fields specify an array of
416 * - LNET_MD_MAX_SIZE: The max_size field is valid.
419 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
420 * capability for memory descriptors. They can't be both set.
421 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
422 * region (i.e. sum of all fragment lengths) must not be less than
425 unsigned int options;
427 * A user-specified value that is associated with the memory
428 * descriptor. The value does not need to be a pointer, but must fit
429 * in the space used by a pointer. This value is recorded in events
430 * associated with operations on this MD.
434 * A handle for the event queue used to log the operations performed on
435 * the memory region. If this argument is a NULL handle (i.e. nullified
436 * by LNetInvalidateHandle()), operations performed on this memory
437 * descriptor are not logged.
439 lnet_handle_eq_t eq_handle;
442 /* Max Transfer Unit (minimum supported everywhere).
443 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
444 * these limits are system wide and not interface-local. */
445 #define LNET_MTU_BITS 20
446 #define LNET_MTU (1 << LNET_MTU_BITS)
448 /** limit on the number of fragments in discontiguous MDs */
449 #define LNET_MAX_IOV 256
452 * Options for the MD structure. See lnet_md_t::options.
454 #define LNET_MD_OP_PUT (1 << 0)
455 /** See lnet_md_t::options. */
456 #define LNET_MD_OP_GET (1 << 1)
457 /** See lnet_md_t::options. */
458 #define LNET_MD_MANAGE_REMOTE (1 << 2)
459 /* unused (1 << 3) */
460 /** See lnet_md_t::options. */
461 #define LNET_MD_TRUNCATE (1 << 4)
462 /** See lnet_md_t::options. */
463 #define LNET_MD_ACK_DISABLE (1 << 5)
464 /** See lnet_md_t::options. */
465 #define LNET_MD_IOVEC (1 << 6)
466 /** See lnet_md_t::options. */
467 #define LNET_MD_MAX_SIZE (1 << 7)
468 /** See lnet_md_t::options. */
469 #define LNET_MD_KIOV (1 << 8)
471 /* For compatibility with Cray Portals */
472 #define LNET_MD_PHYS 0
474 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
475 #define LNET_MD_THRESH_INF (-1)
477 /* NB lustre portals uses struct iovec internally! */
478 typedef struct iovec lnet_md_iovec_t;
481 * A page-based fragment of a MD.
484 /** Pointer to the page where the fragment resides */
485 struct page *kiov_page;
486 /** Length in bytes of the fragment */
487 unsigned int kiov_len;
489 * Starting offset of the fragment within the page. Note that the
490 * end of the fragment must not pass the end of the page; i.e.,
491 * kiov_len + kiov_offset <= PAGE_CACHE_SIZE.
493 unsigned int kiov_offset;
497 /** \addtogroup lnet_eq
501 * Six types of events can be logged in an event queue.
504 /** An incoming GET operation has completed on the MD. */
507 * An incoming PUT operation has completed on the MD. The
508 * underlying layers will not alter the memory (on behalf of this
509 * operation) once this event has been logged.
513 * A REPLY operation has completed. This event is logged after the
514 * data (if any) from the REPLY has been written into the MD.
517 /** An acknowledgment has been received. */
520 * An outgoing send (PUT or GET) operation has completed. This event
521 * is logged after the entire buffer has been sent and it is safe for
522 * the caller to reuse the buffer.
525 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
526 * happen even when the message has not yet been put out on wire.
527 * - It's unsafe to assume that in an outgoing GET operation
528 * the LNET_EVENT_SEND event would happen before the
529 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
530 * LNET_EVENT_ACK events in an outgoing PUT operation.
534 * A MD has been unlinked. Note that LNetMDUnlink() does not
535 * necessarily trigger an LNET_EVENT_UNLINK event.
541 #define LNET_SEQ_BASETYPE long
542 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
543 #define LNET_SEQ_GT(a,b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
546 * Information about an event on a MD.
549 /** The identifier (nid, pid) of the target. */
550 lnet_process_id_t target;
551 /** The identifier (nid, pid) of the initiator. */
552 lnet_process_id_t initiator;
554 * The NID of the immediate sender. If the request has been forwarded
555 * by routers, this is the NID of the last hop; otherwise it's the
556 * same as the initiator.
559 /** Indicates the type of the event. */
560 lnet_event_kind_t type;
561 /** The portal table index specified in the request */
562 unsigned int pt_index;
563 /** A copy of the match bits specified in the request. */
565 /** The length (in bytes) specified in the request. */
566 unsigned int rlength;
568 * The length (in bytes) of the data that was manipulated by the
569 * operation. For truncated operations, the manipulated length will be
570 * the number of bytes specified by the MD (possibly with an offset,
571 * see lnet_md_t). For all other operations, the manipulated length
572 * will be the length of the requested operation, i.e. rlength.
574 unsigned int mlength;
576 * The handle to the MD associated with the event. The handle may be
577 * invalid if the MD has been unlinked.
579 lnet_handle_md_t md_handle;
581 * A snapshot of the state of the MD immediately after the event has
582 * been processed. In particular, the threshold field in md will
583 * reflect the value of the threshold after the operation occurred.
587 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
592 * Indicates the completion status of the operation. It's 0 for
593 * successful operations, otherwise it's an error code.
597 * Indicates whether the MD has been unlinked. Note that:
598 * - An event with unlinked set is the last event on the MD.
599 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
604 * The displacement (in bytes) into the memory region that the
605 * operation used. The offset can be determined by the operation for
606 * a remote managed MD or by the local MD.
607 * \see lnet_md_t::options
611 * The sequence number for this event. Sequence numbers are unique
614 volatile lnet_seq_t sequence;
618 * Event queue handler function type.
620 * The EQ handler runs for each event that is deposited into the EQ. The
621 * handler is supplied with a pointer to the event that triggered the
622 * handler invocation.
624 * The handler must not block, must be reentrant, and must not call any LNet
625 * API functions. It should return as quickly as possible.
627 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
628 #define LNET_EQ_HANDLER_NONE NULL
631 /** \addtogroup lnet_data
635 * Specify whether an acknowledgment should be sent by target when the PUT
636 * operation completes (i.e., when the data has been written to a MD of the
639 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
640 * acknowledgments can be disabled for a MD.
643 /** Request an acknowledgment */
645 /** Request that no acknowledgment should be generated. */