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/
32 #ifndef __UAPI_LNET_TYPES_H__
33 #define __UAPI_LNET_TYPES_H__
38 #include <linux/lnet/lnet-idl.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
50 /** wildcard NID that matches any end-point address */
51 #define LNET_NID_ANY ((lnet_nid_t) -1)
52 /** wildcard PID that matches any lnet_pid_t */
53 #define LNET_PID_ANY ((lnet_pid_t) -1)
55 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
56 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
57 #define LNET_PID_LUSTRE 12345
59 /* how an LNET NID encodes net:address */
60 /** extract the address part of an lnet_nid_t */
62 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
64 return nid & 0xffffffff;
67 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
69 return (nid >> 32) & 0xffffffff;
72 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
74 return (((__u64)net) << 32) | addr;
77 static inline __u32 LNET_NETNUM(__u32 net)
82 static inline __u32 LNET_NETTYP(__u32 net)
84 return (net >> 16) & 0xffff;
87 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
89 return (type << 16) | num;
92 /** The lolnd NID (i.e. myself) */
93 #define LNET_NID_LO_0 LNET_MKNID(LNET_MKNET(LOLND, 0), 0)
95 #define LNET_NET_ANY LNET_NIDNET(LNET_NID_ANY)
97 struct lnet_counters_health {
99 __u32 lch_resend_count;
100 __u32 lch_response_timeout_count;
101 __u32 lch_local_interrupt_count;
102 __u32 lch_local_dropped_count;
103 __u32 lch_local_aborted_count;
104 __u32 lch_local_no_route_count;
105 __u32 lch_local_timeout_count;
106 __u32 lch_local_error_count;
107 __u32 lch_remote_dropped_count;
108 __u32 lch_remote_error_count;
109 __u32 lch_remote_timeout_count;
110 __u32 lch_network_timeout_count;
113 struct lnet_counters {
114 struct lnet_counters_common lct_common;
115 struct lnet_counters_health lct_health;
119 * This is a hard-coded limit on the number of interfaces supported by
120 * the interface bonding implemented by the ksocknal LND. It must be
121 * defined here because it is used in LNet data structures that are
122 * common to all LNDs.
124 #define LNET_INTERFACES_NUM 16
126 /* The minimum number of interfaces per node supported by LNet. */
127 #define LNET_INTERFACES_MIN 16
128 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
129 #define LNET_INTERFACES_MAX_DEFAULT 200
132 * Objects maintained by the LNet are accessed through handles. Handle types
133 * have names of the form lnet_handle_xx, where xx is one of the two letter
134 * object type codes ('md' for memory descriptor, and
135 * 'me' for match entry). Each type of object is given a unique handle type
136 * to enhance type checking.
138 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
140 struct lnet_handle_md {
145 * Invalidate md handle \a h.
147 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
149 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
153 * Check whether handler \a h is invalid.
155 * \return 1 if handle is invalid, 0 if valid.
157 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
159 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
165 struct lnet_process_id {
173 /** \addtogroup lnet_me
177 * Specifies whether the match entry or memory descriptor should be unlinked
178 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
186 * Values of the type enum lnet_ins_pos are used to control where a new match
187 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
188 * entry before the current entry or before the head of the list. The value
189 * LNET_INS_AFTER is used to insert the new entry after the current entry
190 * or after the last item in the list.
193 /** insert ME before current position or head of the list */
195 /** insert ME after current position or tail of the list */
197 /** attach ME at tail of local CPU partition ME list */
203 /** \addtogroup lnet_md
207 * Event queue handler function type.
209 * The EQ handler runs for each event that is deposited into the EQ. The
210 * handler is supplied with a pointer to the event that triggered the
211 * handler invocation.
213 * The handler must not block, must be reentrant, and must not call any LNet
214 * API functions. It should return as quickly as possible.
217 typedef void (*lnet_handler_t)(struct lnet_event *event);
220 * Defines the visible parts of a memory descriptor. Values of this type
221 * are used to initialize memory descriptors.
225 * Specify the memory region associated with the memory descriptor.
226 * If the options field has:
227 * - LNET_MD_KIOV bit set: The start field points to the starting
228 * address of an array of struct bio_vec and the length field specifies
229 * the number of entries in the array. The length can't be bigger
230 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
231 * fragments that are not necessarily mapped in virtal memory.
232 * - Otherwise: The memory region is contiguous. The start field
233 * specifies the starting address for the memory region and the
234 * length field specifies its length.
236 * When the memory region is fragmented, all fragments but the first
237 * one must start on page boundary, and all but the last must end on
243 * Specifies the maximum number of operations that can be performed
244 * on the memory descriptor. An operation is any action that could
245 * possibly generate an event. In the usual case, the threshold value
246 * is decremented for each operation on the MD. When the threshold
247 * drops to zero, the MD becomes inactive and does not respond to
248 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
249 * there is no bound on the number of operations that may be applied
254 * Specifies the largest incoming request that the memory descriptor
255 * should respond to. When the unused portion of a MD (length -
256 * local offset) falls below this value, the MD becomes inactive and
257 * does not respond to further operations. This value is only used
258 * if the LNET_MD_MAX_SIZE option is set.
262 * Specifies the behavior of the memory descriptor. A bitwise OR
263 * of the following values can be used:
264 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
265 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
266 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
267 * region is provided by the incoming request. By default, the
268 * offset is maintained locally. When maintained locally, the
269 * offset is incremented by the length of the request so that
270 * the next operation (PUT or GET) will access the next part of
271 * the memory region. Note that only one offset variable exists
272 * per memory descriptor. If both PUT and GET operations are
273 * performed on a memory descriptor, the offset is updated each time.
274 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
275 * be reduced to match the memory available in the region (determined
276 * by subtracting the offset from the length of the memory region).
277 * By default, if the length in the incoming operation is greater
278 * than the amount of memory available, the operation is rejected.
279 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
280 * incoming PUT operations, even if requested. By default,
281 * acknowledgments are sent for PUT operations that request an
282 * acknowledgment. Acknowledgments are never sent for GET operations.
283 * The data sent in the REPLY serves as an implicit acknowledgment.
284 * - LNET_MD_KIOV: The start and length fields specify an array of
286 * - LNET_MD_MAX_SIZE: The max_size field is valid.
287 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
288 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
289 * regardless of the value of the lnet_response_tracking param.
290 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
291 * regardless of the value of the lnet_response_tracking param.
294 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
296 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
297 * region (i.e. sum of all fragment lengths) must not be less than
300 unsigned int options;
302 * A user-specified value that is associated with the memory
303 * descriptor. The value does not need to be a pointer, but must fit
304 * in the space used by a pointer. This value is recorded in events
305 * associated with operations on this MD.
309 * The event handler used to log the operations performed on
310 * the memory region. If this argument is NULL operations
311 * performed on this memory descriptor are not logged.
313 lnet_handler_t handler;
315 * The bulk MD handle which was registered to describe the buffers
316 * either to be used to transfer data to the peer or receive data
317 * from the peer. This allows LNet to properly determine the NUMA
318 * node on which the memory was allocated and use that to select the
319 * nearest local network interface. This value is only used
320 * if the LNET_MD_BULK_HANDLE option is set.
322 struct lnet_handle_md bulk_handle;
325 /* Max Transfer Unit (minimum supported everywhere).
326 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
327 * these limits are system wide and not interface-local. */
328 #define LNET_MTU_BITS 20
329 #define LNET_MTU (1 << LNET_MTU_BITS)
332 * Options for the MD structure. See struct lnet_md::options.
334 #define LNET_MD_OP_PUT (1 << 0)
335 /** See struct lnet_md::options. */
336 #define LNET_MD_OP_GET (1 << 1)
337 /** See struct lnet_md::options. */
338 #define LNET_MD_MANAGE_REMOTE (1 << 2)
339 /* unused (1 << 3) */
340 /** See struct lnet_md::options. */
341 #define LNET_MD_TRUNCATE (1 << 4)
342 /** See struct lnet_md::options. */
343 #define LNET_MD_ACK_DISABLE (1 << 5)
344 /** See struct lnet_md::options. */
345 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
346 /** See struct lnet_md::options. */
347 #define LNET_MD_MAX_SIZE (1 << 7)
348 /** See struct lnet_md::options. */
349 #define LNET_MD_KIOV (1 << 8)
350 /** See struct lnet_md::options. */
351 #define LNET_MD_BULK_HANDLE (1 << 9)
352 /** See struct lnet_md::options. */
353 #define LNET_MD_TRACK_RESPONSE (1 << 10)
354 /** See struct lnet_md::options. */
355 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
357 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
358 #define LNET_MD_THRESH_INF (-1)
362 /** \addtogroup lnet_eq
366 * Six types of events can be logged in an event queue.
368 enum lnet_event_kind {
369 /** An incoming GET operation has completed on the MD. */
372 * An incoming PUT operation has completed on the MD. The
373 * underlying layers will not alter the memory (on behalf of this
374 * operation) once this event has been logged.
378 * A REPLY operation has completed. This event is logged after the
379 * data (if any) from the REPLY has been written into the MD.
382 /** An acknowledgment has been received. */
385 * An outgoing send (PUT or GET) operation has completed. This event
386 * is logged after the entire buffer has been sent and it is safe for
387 * the caller to reuse the buffer.
390 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
391 * happen even when the message has not yet been put out on wire.
392 * - It's unsafe to assume that in an outgoing GET operation
393 * the LNET_EVENT_SEND event would happen before the
394 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
395 * LNET_EVENT_ACK events in an outgoing PUT operation.
399 * A MD has been unlinked. Note that LNetMDUnlink() does not
400 * necessarily trigger an LNET_EVENT_UNLINK event.
406 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
409 * Information about an event on a MD.
412 /** The identifier (nid, pid) of the target. */
413 struct lnet_process_id target;
414 /** The identifier (nid, pid) of the initiator. */
415 struct lnet_process_id initiator;
416 /** The source NID on the initiator. */
417 struct lnet_process_id source;
419 * The NID of the immediate sender. If the request has been forwarded
420 * by routers, this is the NID of the last hop; otherwise it's the
421 * same as the source.
424 /** Indicates the type of the event. */
425 enum lnet_event_kind type;
426 /** The portal table index specified in the request */
427 unsigned int pt_index;
428 /** A copy of the match bits specified in the request. */
430 /** The length (in bytes) specified in the request. */
431 unsigned int rlength;
433 * The length (in bytes) of the data that was manipulated by the
434 * operation. For truncated operations, the manipulated length will be
435 * the number of bytes specified by the MD (possibly with an offset,
436 * see struct lnet_md). For all other operations, the manipulated length
437 * will be the length of the requested operation, i.e. rlength.
439 unsigned int mlength;
441 * The handle to the MD associated with the event. The handle may be
442 * invalid if the MD has been unlinked.
444 struct lnet_handle_md md_handle;
446 * A snapshot of relevant state of the MD immediately after the event
447 * has been processed.
451 unsigned int md_options;
453 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
458 * The message type, to ensure a handler for LNET_EVENT_SEND can
459 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
463 * Indicates the completion status of the operation. It's 0 for
464 * successful operations, otherwise it's an error code.
468 * Indicates whether the MD has been unlinked. Note that:
469 * - An event with unlinked set is the last event on the MD.
470 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
475 * The displacement (in bytes) into the memory region that the
476 * operation used. The offset can be determined by the operation for
477 * a remote managed MD or by the local MD.
478 * \see struct lnet_md::options
482 * The sequence number for this event. Sequence numbers are unique
485 volatile unsigned long sequence;
488 /** \addtogroup lnet_data
492 * Specify whether an acknowledgment should be sent by target when the PUT
493 * operation completes (i.e., when the data has been written to a MD of the
496 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
497 * by which acknowledgments can be disabled for a MD.
500 /** Request an acknowledgment */
502 /** Request that no acknowledgment should be generated. */
507 * UDSP action types. There are two available actions:
508 * 1. PRIORITY - set priority of matching LNet constructs
509 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
511 enum lnet_udsp_action_type {
512 EN_LNET_UDSP_ACTION_NONE = 0,
513 /** assign a priority to matching constructs */
514 EN_LNET_UDSP_ACTION_PRIORITY = 1,
515 /** assign a preferred list of NIDs to matching constructs */
516 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,