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>
41 /** \addtogroup lnet_addr
44 /** Portal reserved for LNet's own use.
45 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
47 #define LNET_RESERVED_PORTAL 0
50 * Address of an end-point in an LNet network.
52 * A node can have multiple end-points and hence multiple addresses.
53 * An LNet network can be a simple network (e.g. tcp0) or a network of
54 * LNet networks connected by LNet routers. Therefore an end-point address
55 * has two parts: network ID, and address within a network.
57 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
59 typedef __u64 lnet_nid_t;
61 * ID of a process in a node. Shortened as PID to distinguish from
62 * lnet_process_id_t, the global process ID.
64 typedef __u32 lnet_pid_t;
66 /** wildcard NID that matches any end-point address */
67 #define LNET_NID_ANY ((lnet_nid_t) -1)
68 /** wildcard PID that matches any lnet_pid_t */
69 #define LNET_PID_ANY ((lnet_pid_t) -1)
71 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
72 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
73 #define LNET_PID_LUSTRE 12345
75 #define LNET_TIME_FOREVER (-1)
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 __packed
112 /* Packed version of lnet_process_id_t to transfer via network */
114 /* node id / process id */
117 } WIRE_ATTR lnet_process_id_packed_t;
119 /* The wire handle's interface cookie only matches one network interface in
120 * one epoch (i.e. new cookie when the interface restarts or the node
121 * reboots). The object cookie only matches one object on that interface
122 * during that object's lifetime (i.e. no cookie re-use). */
124 __u64 wh_interface_cookie;
125 __u64 wh_object_cookie;
126 } WIRE_ATTR lnet_handle_wire_t;
136 /* The variant fields of the portals message header are aligned on an 8
137 * byte boundary in the message header. Note that all types used in these
138 * wire structs MUST be fixed size and the smaller types are placed at the
140 typedef struct lnet_ack {
141 lnet_handle_wire_t dst_wmd;
144 } WIRE_ATTR lnet_ack_t;
146 typedef struct lnet_put {
147 lnet_handle_wire_t ack_wmd;
152 } WIRE_ATTR lnet_put_t;
154 typedef struct lnet_get {
155 lnet_handle_wire_t return_wmd;
160 } WIRE_ATTR lnet_get_t;
162 typedef struct lnet_reply {
163 lnet_handle_wire_t dst_wmd;
164 } WIRE_ATTR lnet_reply_t;
166 typedef struct lnet_hello {
169 } WIRE_ATTR lnet_hello_t;
176 __u32 type; /* lnet_msg_type_t */
177 __u32 payload_length; /* payload data to follow */
178 /*<------__u64 aligned------->*/
186 } WIRE_ATTR lnet_hdr_t;
188 /* A HELLO message contains a magic number and protocol version
189 * code in the header's dest_nid, the peer's NID in the src_nid, and
190 * LNET_MSG_HELLO in the type field. All other common fields are zero
191 * (including payload_size; i.e. no payload).
192 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
193 * running the same protocol and to find out its NID. These LNDs should
194 * exchange HELLO messages when a connection is first established. Individual
195 * LNDs can put whatever else they fancy in lnet_hdr_t::msg.
198 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
199 __u16 version_major; /* increment on incompatible change */
200 __u16 version_minor; /* increment on compatible change */
201 } WIRE_ATTR lnet_magicversion_t;
203 /* PROTO MAGIC for LNDs */
204 #define LNET_PROTO_IB_MAGIC 0x0be91b91
205 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
206 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
207 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
208 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
210 /* Placeholder for a future "unified" protocol across all LNDs */
211 /* Current LNDs that receive a request with this magic will respond with a
212 * "stub" reply using their current protocol */
213 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
215 #define LNET_PROTO_TCP_VERSION_MAJOR 1
216 #define LNET_PROTO_TCP_VERSION_MINOR 0
218 /* Acceptor connection request */
220 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
221 __u32 acr_version; /* protocol version */
222 __u64 acr_nid; /* target NID */
223 } WIRE_ATTR lnet_acceptor_connreq_t;
225 #define LNET_PROTO_ACCEPTOR_VERSION 1
231 } WIRE_ATTR lnet_ni_status_t;
238 lnet_ni_status_t pi_ni[0];
239 } WIRE_ATTR lnet_ping_info_t;
241 typedef struct lnet_counters {
253 } WIRE_ATTR lnet_counters_t;
255 #define LNET_NI_STATUS_UP 0x15aac0de
256 #define LNET_NI_STATUS_DOWN 0xdeadface
257 #define LNET_NI_STATUS_INVALID 0x00000000
259 #define LNET_MAX_INTERFACES 16
262 * Objects maintained by the LNet are accessed through handles. Handle types
263 * have names of the form lnet_handle_xx_t, where xx is one of the two letter
264 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
265 * 'me' for match entry).
266 * Each type of object is given a unique handle type to enhance type checking.
267 * The type lnet_handle_any_t can be used when a generic handle is needed.
268 * Every handle value can be converted into a value of type lnet_handle_any_t
269 * without loss of information.
275 typedef lnet_handle_any_t lnet_handle_eq_t;
276 typedef lnet_handle_any_t lnet_handle_md_t;
277 typedef lnet_handle_any_t lnet_handle_me_t;
279 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
282 * Invalidate handle \a h.
284 static inline void LNetInvalidateHandle(lnet_handle_any_t *h)
286 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
290 * Compare handles \a h1 and \a h2.
292 * \return 1 if handles are equal, 0 if otherwise.
294 static inline int LNetHandleIsEqual(lnet_handle_any_t h1, lnet_handle_any_t h2)
296 return h1.cookie == h2.cookie;
300 * Check whether handle \a h is invalid.
302 * \return 1 if handle is invalid, 0 if valid.
304 static inline int LNetHandleIsInvalid(lnet_handle_any_t h)
306 return h.cookie == LNET_WIRE_HANDLE_COOKIE_NONE;
320 /** \addtogroup lnet_me
324 * Specifies whether the match entry or memory descriptor should be unlinked
325 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
333 * Values of the type lnet_ins_pos_t are used to control where a new match
334 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
335 * entry before the current entry or before the head of the list. The value
336 * LNET_INS_AFTER is used to insert the new entry after the current entry
337 * or after the last item in the list.
340 /** insert ME before current position or head of the list */
342 /** insert ME after current position or tail of the list */
344 /** attach ME at tail of local CPU partition ME list */
350 /** \addtogroup lnet_md
354 * Defines the visible parts of a memory descriptor. Values of this type
355 * are used to initialize memory descriptors.
359 * Specify the memory region associated with the memory descriptor.
360 * If the options field has:
361 * - LNET_MD_KIOV bit set: The start field points to the starting
362 * address of an array of lnet_kiov_t and the length field specifies
363 * the number of entries in the array. The length can't be bigger
364 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
365 * fragments that are not necessarily mapped in virtual memory.
366 * - LNET_MD_IOVEC bit set: The start field points to the starting
367 * address of an array of struct iovec 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 struct iovec is used to describe fragments
370 * that have virtual addresses.
371 * - Otherwise: The memory region is contiguous. The start field
372 * specifies the starting address for the memory region and the
373 * length field specifies its length.
375 * When the memory region is fragmented, all fragments but the first
376 * one must start on page boundary, and all but the last must end on
382 * Specifies the maximum number of operations that can be performed
383 * on the memory descriptor. An operation is any action that could
384 * possibly generate an event. In the usual case, the threshold value
385 * is decremented for each operation on the MD. When the threshold
386 * drops to zero, the MD becomes inactive and does not respond to
387 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
388 * there is no bound on the number of operations that may be applied
393 * Specifies the largest incoming request that the memory descriptor
394 * should respond to. When the unused portion of a MD (length -
395 * local offset) falls below this value, the MD becomes inactive and
396 * does not respond to further operations. This value is only used
397 * if the LNET_MD_MAX_SIZE option is set.
401 * Specifies the behavior of the memory descriptor. A bitwise OR
402 * of the following values can be used:
403 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
404 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
405 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
406 * region is provided by the incoming request. By default, the
407 * offset is maintained locally. When maintained locally, the
408 * offset is incremented by the length of the request so that
409 * the next operation (PUT or GET) will access the next part of
410 * the memory region. Note that only one offset variable exists
411 * per memory descriptor. If both PUT and GET operations are
412 * performed on a memory descriptor, the offset is updated each time.
413 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
414 * be reduced to match the memory available in the region (determined
415 * by subtracting the offset from the length of the memory region).
416 * By default, if the length in the incoming operation is greater
417 * than the amount of memory available, the operation is rejected.
418 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
419 * incoming PUT operations, even if requested. By default,
420 * acknowledgments are sent for PUT operations that request an
421 * acknowledgment. Acknowledgments are never sent for GET operations.
422 * The data sent in the REPLY serves as an implicit acknowledgment.
423 * - LNET_MD_KIOV: The start and length fields specify an array of
425 * - LNET_MD_IOVEC: The start and length fields specify an array of
427 * - LNET_MD_MAX_SIZE: The max_size field is valid.
430 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
431 * capability for memory descriptors. They can't be both set.
432 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
433 * region (i.e. sum of all fragment lengths) must not be less than
436 unsigned int options;
438 * A user-specified value that is associated with the memory
439 * descriptor. The value does not need to be a pointer, but must fit
440 * in the space used by a pointer. This value is recorded in events
441 * associated with operations on this MD.
445 * A handle for the event queue used to log the operations performed on
446 * the memory region. If this argument is a NULL handle (i.e. nullified
447 * by LNetInvalidateHandle()), operations performed on this memory
448 * descriptor are not logged.
450 lnet_handle_eq_t eq_handle;
453 /* Max Transfer Unit (minimum supported everywhere).
454 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
455 * these limits are system wide and not interface-local. */
456 #define LNET_MTU_BITS 20
457 #define LNET_MTU (1 << LNET_MTU_BITS)
459 /** limit on the number of fragments in discontiguous MDs */
460 #define LNET_MAX_IOV 256
463 * Options for the MD structure. See lnet_md_t::options.
465 #define LNET_MD_OP_PUT (1 << 0)
466 /** See lnet_md_t::options. */
467 #define LNET_MD_OP_GET (1 << 1)
468 /** See lnet_md_t::options. */
469 #define LNET_MD_MANAGE_REMOTE (1 << 2)
470 /* unused (1 << 3) */
471 /** See lnet_md_t::options. */
472 #define LNET_MD_TRUNCATE (1 << 4)
473 /** See lnet_md_t::options. */
474 #define LNET_MD_ACK_DISABLE (1 << 5)
475 /** See lnet_md_t::options. */
476 #define LNET_MD_IOVEC (1 << 6)
477 /** See lnet_md_t::options. */
478 #define LNET_MD_MAX_SIZE (1 << 7)
479 /** See lnet_md_t::options. */
480 #define LNET_MD_KIOV (1 << 8)
482 /* For compatibility with Cray Portals */
483 #define LNET_MD_PHYS 0
485 /** Infinite threshold on MD operations. See lnet_md_t::threshold */
486 #define LNET_MD_THRESH_INF (-1)
488 /* NB lustre portals uses struct iovec internally! */
489 typedef struct iovec lnet_md_iovec_t;
492 * A page-based fragment of a MD.
495 /** Pointer to the page where the fragment resides */
496 struct page *kiov_page;
497 /** Length in bytes of the fragment */
498 unsigned int kiov_len;
500 * Starting offset of the fragment within the page. Note that the
501 * end of the fragment must not pass the end of the page; i.e.,
502 * kiov_len + kiov_offset <= PAGE_CACHE_SIZE.
504 unsigned int kiov_offset;
508 /** \addtogroup lnet_eq
512 * Six types of events can be logged in an event queue.
515 /** An incoming GET operation has completed on the MD. */
518 * An incoming PUT operation has completed on the MD. The
519 * underlying layers will not alter the memory (on behalf of this
520 * operation) once this event has been logged.
524 * A REPLY operation has completed. This event is logged after the
525 * data (if any) from the REPLY has been written into the MD.
528 /** An acknowledgment has been received. */
531 * An outgoing send (PUT or GET) operation has completed. This event
532 * is logged after the entire buffer has been sent and it is safe for
533 * the caller to reuse the buffer.
536 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
537 * happen even when the message has not yet been put out on wire.
538 * - It's unsafe to assume that in an outgoing GET operation
539 * the LNET_EVENT_SEND event would happen before the
540 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
541 * LNET_EVENT_ACK events in an outgoing PUT operation.
545 * A MD has been unlinked. Note that LNetMDUnlink() does not
546 * necessarily trigger an LNET_EVENT_UNLINK event.
552 #define LNET_SEQ_BASETYPE long
553 typedef unsigned LNET_SEQ_BASETYPE lnet_seq_t;
554 #define LNET_SEQ_GT(a, b) (((signed LNET_SEQ_BASETYPE)((a) - (b))) > 0)
557 * Information about an event on a MD.
560 /** The identifier (nid, pid) of the target. */
561 lnet_process_id_t target;
562 /** The identifier (nid, pid) of the initiator. */
563 lnet_process_id_t initiator;
565 * The NID of the immediate sender. If the request has been forwarded
566 * by routers, this is the NID of the last hop; otherwise it's the
567 * same as the initiator.
570 /** Indicates the type of the event. */
571 lnet_event_kind_t type;
572 /** The portal table index specified in the request */
573 unsigned int pt_index;
574 /** A copy of the match bits specified in the request. */
576 /** The length (in bytes) specified in the request. */
577 unsigned int rlength;
579 * The length (in bytes) of the data that was manipulated by the
580 * operation. For truncated operations, the manipulated length will be
581 * the number of bytes specified by the MD (possibly with an offset,
582 * see lnet_md_t). For all other operations, the manipulated length
583 * will be the length of the requested operation, i.e. rlength.
585 unsigned int mlength;
587 * The handle to the MD associated with the event. The handle may be
588 * invalid if the MD has been unlinked.
590 lnet_handle_md_t md_handle;
592 * A snapshot of the state of the MD immediately after the event has
593 * been processed. In particular, the threshold field in md will
594 * reflect the value of the threshold after the operation occurred.
598 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
603 * Indicates the completion status of the operation. It's 0 for
604 * successful operations, otherwise it's an error code.
608 * Indicates whether the MD has been unlinked. Note that:
609 * - An event with unlinked set is the last event on the MD.
610 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
615 * The displacement (in bytes) into the memory region that the
616 * operation used. The offset can be determined by the operation for
617 * a remote managed MD or by the local MD.
618 * \see lnet_md_t::options
622 * The sequence number for this event. Sequence numbers are unique
625 volatile lnet_seq_t sequence;
629 * Event queue handler function type.
631 * The EQ handler runs for each event that is deposited into the EQ. The
632 * handler is supplied with a pointer to the event that triggered the
633 * handler invocation.
635 * The handler must not block, must be reentrant, and must not call any LNet
636 * API functions. It should return as quickly as possible.
638 typedef void (*lnet_eq_handler_t)(lnet_event_t *event);
639 #define LNET_EQ_HANDLER_NONE NULL
642 /** \addtogroup lnet_data
646 * Specify whether an acknowledgment should be sent by target when the PUT
647 * operation completes (i.e., when the data has been written to a MD of the
650 * \see lnet_md_t::options for the discussion on LNET_MD_ACK_DISABLE by which
651 * acknowledgments can be disabled for a MD.
654 /** Request an acknowledgment */
656 /** Request that no acknowledgment should be generated. */