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, 2017, 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/lnet/lnet-idl.h>
41 /** \addtogroup lnet_addr
44 #define LNET_VERSION "0.7.0"
46 /** Portal reserved for LNet's own use.
47 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
49 #define LNET_RESERVED_PORTAL 0
51 /** wildcard NID that matches any end-point address */
52 #define LNET_NID_ANY ((lnet_nid_t) -1)
53 /** wildcard PID that matches any lnet_pid_t */
54 #define LNET_PID_ANY ((lnet_pid_t) -1)
56 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
57 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
58 #define LNET_PID_LUSTRE 12345
60 /* how an LNET NID encodes net:address */
61 /** extract the address part of an lnet_nid_t */
63 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
65 return nid & 0xffffffff;
68 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
70 return (nid >> 32) & 0xffffffff;
73 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
75 return (((__u64)net) << 32) | addr;
78 static inline __u32 LNET_NETNUM(__u32 net)
83 static inline __u32 LNET_NETTYP(__u32 net)
85 return (net >> 16) & 0xffff;
88 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
90 return (type << 16) | num;
93 /** The lolnd NID (i.e. myself) */
94 #define LNET_NID_LO_0 LNET_MKNID(LNET_MKNET(LOLND, 0), 0)
96 #define LNET_NET_ANY LNET_NIDNET(LNET_NID_ANY)
98 struct lnet_counters_health {
100 __u32 lch_resend_count;
101 __u32 lch_response_timeout_count;
102 __u32 lch_local_interrupt_count;
103 __u32 lch_local_dropped_count;
104 __u32 lch_local_aborted_count;
105 __u32 lch_local_no_route_count;
106 __u32 lch_local_timeout_count;
107 __u32 lch_local_error_count;
108 __u32 lch_remote_dropped_count;
109 __u32 lch_remote_error_count;
110 __u32 lch_remote_timeout_count;
111 __u32 lch_network_timeout_count;
114 struct lnet_counters {
115 struct lnet_counters_common lct_common;
116 struct lnet_counters_health lct_health;
120 * This is a hard-coded limit on the number of interfaces supported by
121 * the interface bonding implemented by the ksocknal LND. It must be
122 * defined here because it is used in LNet data structures that are
123 * common to all LNDs.
125 #define LNET_INTERFACES_NUM 16
127 /* The minimum number of interfaces per node supported by LNet. */
128 #define LNET_INTERFACES_MIN 16
129 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
130 #define LNET_INTERFACES_MAX_DEFAULT 200
133 * Objects maintained by the LNet are accessed through handles. Handle types
134 * have names of the form lnet_handle_xx, where xx is one of the two letter
135 * object type codes ('md' for memory descriptor, and
136 * 'me' for match entry). Each type of object is given a unique handle type
137 * to enhance type checking.
139 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
141 struct lnet_handle_md {
146 * Invalidate md handle \a h.
148 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
150 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
154 * Check whether handler \a h is invalid.
156 * \return 1 if handle is invalid, 0 if valid.
158 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
160 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
166 struct lnet_process_id {
174 /** \addtogroup lnet_me
178 * Specifies whether the match entry or memory descriptor should be unlinked
179 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
187 * Values of the type enum lnet_ins_pos are used to control where a new match
188 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
189 * entry before the current entry or before the head of the list. The value
190 * LNET_INS_AFTER is used to insert the new entry after the current entry
191 * or after the last item in the list.
194 /** insert ME before current position or head of the list */
196 /** insert ME after current position or tail of the list */
198 /** attach ME at tail of local CPU partition ME list */
204 /** \addtogroup lnet_md
208 * Event queue handler function type.
210 * The EQ handler runs for each event that is deposited into the EQ. The
211 * handler is supplied with a pointer to the event that triggered the
212 * handler invocation.
214 * The handler must not block, must be reentrant, and must not call any LNet
215 * API functions. It should return as quickly as possible.
218 typedef void (*lnet_handler_t)(struct lnet_event *event);
221 * Defines the visible parts of a memory descriptor. Values of this type
222 * are used to initialize memory descriptors.
226 * Specify the memory region associated with the memory descriptor.
227 * If the options field has:
228 * - LNET_MD_KIOV bit set: The start field points to the starting
229 * address of an array of struct bio_vec and the length field specifies
230 * the number of entries in the array. The length can't be bigger
231 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
232 * fragments that are not necessarily mapped in virtal memory.
233 * - Otherwise: The memory region is contiguous. The start field
234 * specifies the starting address for the memory region and the
235 * length field specifies its length.
237 * When the memory region is fragmented, all fragments but the first
238 * one must start on page boundary, and all but the last must end on
244 * Specifies the maximum number of operations that can be performed
245 * on the memory descriptor. An operation is any action that could
246 * possibly generate an event. In the usual case, the threshold value
247 * is decremented for each operation on the MD. When the threshold
248 * drops to zero, the MD becomes inactive and does not respond to
249 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
250 * there is no bound on the number of operations that may be applied
255 * Specifies the largest incoming request that the memory descriptor
256 * should respond to. When the unused portion of a MD (length -
257 * local offset) falls below this value, the MD becomes inactive and
258 * does not respond to further operations. This value is only used
259 * if the LNET_MD_MAX_SIZE option is set.
263 * Specifies the behavior of the memory descriptor. A bitwise OR
264 * of the following values can be used:
265 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
266 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
267 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
268 * region is provided by the incoming request. By default, the
269 * offset is maintained locally. When maintained locally, the
270 * offset is incremented by the length of the request so that
271 * the next operation (PUT or GET) will access the next part of
272 * the memory region. Note that only one offset variable exists
273 * per memory descriptor. If both PUT and GET operations are
274 * performed on a memory descriptor, the offset is updated each time.
275 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
276 * be reduced to match the memory available in the region (determined
277 * by subtracting the offset from the length of the memory region).
278 * By default, if the length in the incoming operation is greater
279 * than the amount of memory available, the operation is rejected.
280 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
281 * incoming PUT operations, even if requested. By default,
282 * acknowledgments are sent for PUT operations that request an
283 * acknowledgment. Acknowledgments are never sent for GET operations.
284 * The data sent in the REPLY serves as an implicit acknowledgment.
285 * - LNET_MD_KIOV: The start and length fields specify an array of
287 * - LNET_MD_MAX_SIZE: The max_size field is valid.
288 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
289 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
290 * regardless of the value of the lnet_response_tracking param.
291 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
292 * regardless of the value of the lnet_response_tracking param.
295 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
297 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
298 * region (i.e. sum of all fragment lengths) must not be less than
301 unsigned int options;
303 * A user-specified value that is associated with the memory
304 * descriptor. The value does not need to be a pointer, but must fit
305 * in the space used by a pointer. This value is recorded in events
306 * associated with operations on this MD.
310 * The event handler used to log the operations performed on
311 * the memory region. If this argument is NULL operations
312 * performed on this memory descriptor are not logged.
314 lnet_handler_t handler;
316 * The bulk MD handle which was registered to describe the buffers
317 * either to be used to transfer data to the peer or receive data
318 * from the peer. This allows LNet to properly determine the NUMA
319 * node on which the memory was allocated and use that to select the
320 * nearest local network interface. This value is only used
321 * if the LNET_MD_BULK_HANDLE option is set.
323 struct lnet_handle_md bulk_handle;
326 /* Max Transfer Unit (minimum supported everywhere).
327 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
328 * these limits are system wide and not interface-local. */
329 #define LNET_MTU_BITS 20
330 #define LNET_MTU (1 << LNET_MTU_BITS)
333 * Options for the MD structure. See struct lnet_md::options.
335 #define LNET_MD_OP_PUT (1 << 0)
336 /** See struct lnet_md::options. */
337 #define LNET_MD_OP_GET (1 << 1)
338 /** See struct lnet_md::options. */
339 #define LNET_MD_MANAGE_REMOTE (1 << 2)
340 /* unused (1 << 3) */
341 /** See struct lnet_md::options. */
342 #define LNET_MD_TRUNCATE (1 << 4)
343 /** See struct lnet_md::options. */
344 #define LNET_MD_ACK_DISABLE (1 << 5)
345 /** See struct lnet_md::options. */
346 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
347 /** See struct lnet_md::options. */
348 #define LNET_MD_MAX_SIZE (1 << 7)
349 /** See struct lnet_md::options. */
350 #define LNET_MD_KIOV (1 << 8)
351 /** See struct lnet_md::options. */
352 #define LNET_MD_BULK_HANDLE (1 << 9)
353 /** See struct lnet_md::options. */
354 #define LNET_MD_TRACK_RESPONSE (1 << 10)
355 /** See struct lnet_md::options. */
356 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
358 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
359 #define LNET_MD_THRESH_INF (-1)
363 /** \addtogroup lnet_eq
367 * Six types of events can be logged in an event queue.
369 enum lnet_event_kind {
370 /** An incoming GET operation has completed on the MD. */
373 * An incoming PUT operation has completed on the MD. The
374 * underlying layers will not alter the memory (on behalf of this
375 * operation) once this event has been logged.
379 * A REPLY operation has completed. This event is logged after the
380 * data (if any) from the REPLY has been written into the MD.
383 /** An acknowledgment has been received. */
386 * An outgoing send (PUT or GET) operation has completed. This event
387 * is logged after the entire buffer has been sent and it is safe for
388 * the caller to reuse the buffer.
391 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
392 * happen even when the message has not yet been put out on wire.
393 * - It's unsafe to assume that in an outgoing GET operation
394 * the LNET_EVENT_SEND event would happen before the
395 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
396 * LNET_EVENT_ACK events in an outgoing PUT operation.
400 * A MD has been unlinked. Note that LNetMDUnlink() does not
401 * necessarily trigger an LNET_EVENT_UNLINK event.
407 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
410 * Information about an event on a MD.
413 /** The identifier (nid, pid) of the target. */
414 struct lnet_process_id target;
415 /** The identifier (nid, pid) of the initiator. */
416 struct lnet_process_id initiator;
417 /** The source NID on the initiator. */
418 struct lnet_process_id source;
420 * The NID of the immediate sender. If the request has been forwarded
421 * by routers, this is the NID of the last hop; otherwise it's the
422 * same as the source.
425 /** Indicates the type of the event. */
426 enum lnet_event_kind type;
427 /** The portal table index specified in the request */
428 unsigned int pt_index;
429 /** A copy of the match bits specified in the request. */
431 /** The length (in bytes) specified in the request. */
432 unsigned int rlength;
434 * The length (in bytes) of the data that was manipulated by the
435 * operation. For truncated operations, the manipulated length will be
436 * the number of bytes specified by the MD (possibly with an offset,
437 * see struct lnet_md). For all other operations, the manipulated length
438 * will be the length of the requested operation, i.e. rlength.
440 unsigned int mlength;
442 * The handle to the MD associated with the event. The handle may be
443 * invalid if the MD has been unlinked.
445 struct lnet_handle_md md_handle;
447 * A snapshot of relevant state of the MD immediately after the event
448 * has been processed.
452 unsigned int md_options;
454 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
459 * The message type, to ensure a handler for LNET_EVENT_SEND can
460 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
464 * Indicates the completion status of the operation. It's 0 for
465 * successful operations, otherwise it's an error code.
469 * Indicates whether the MD has been unlinked. Note that:
470 * - An event with unlinked set is the last event on the MD.
471 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
476 * The displacement (in bytes) into the memory region that the
477 * operation used. The offset can be determined by the operation for
478 * a remote managed MD or by the local MD.
479 * \see struct lnet_md::options
483 * The sequence number for this event. Sequence numbers are unique
486 volatile unsigned long sequence;
489 /** \addtogroup lnet_data
493 * Specify whether an acknowledgment should be sent by target when the PUT
494 * operation completes (i.e., when the data has been written to a MD of the
497 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
498 * by which acknowledgments can be disabled for a MD.
501 /** Request an acknowledgment */
503 /** Request that no acknowledgment should be generated. */
508 * UDSP action types. There are two available actions:
509 * 1. PRIORITY - set priority of matching LNet constructs
510 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
512 enum lnet_udsp_action_type {
513 EN_LNET_UDSP_ACTION_NONE = 0,
514 /** assign a priority to matching constructs */
515 EN_LNET_UDSP_ACTION_PRIORITY = 1,
516 /** assign a preferred list of NIDs to matching constructs */
517 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,