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__
35 #include <linux/string.h>
36 #include <asm/byteorder.h>
41 #include <linux/lnet/lnet-idl.h>
43 /** \addtogroup lnet_addr
46 #define LNET_VERSION "0.7.0"
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
53 /** wildcard NID that matches any end-point address */
54 #define LNET_NID_ANY ((lnet_nid_t) -1)
55 /** wildcard PID that matches any lnet_pid_t */
56 #define LNET_PID_ANY ((lnet_pid_t) -1)
58 static inline int LNET_NID_IS_ANY(const struct lnet_nid *nid)
60 /* A NULL pointer can be used to mean "ANY" */
61 return !nid || nid->nid_type == 0xFF;
64 #define LNET_ANY_NID ((struct lnet_nid) \
65 {0xFF, 0xFF, ~0, {~0, ~0, ~0, ~0} })
67 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
68 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
69 #define LNET_PID_LUSTRE 12345
71 /* how an LNET NID encodes net:address */
72 /** extract the address part of an lnet_nid_t */
74 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
76 return nid & 0xffffffff;
79 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
81 return (nid >> 32) & 0xffffffff;
84 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
86 return (((__u64)net) << 32) | addr;
89 static inline __u32 LNET_NETNUM(__u32 net)
94 static inline __u32 LNET_NETTYP(__u32 net)
96 return (net >> 16) & 0xff;
99 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
101 return (type << 16) | num;
104 /** The lolnd NID (i.e. myself) */
105 #define LNET_NID_LO_0 LNET_MKNID(LNET_MKNET(LOLND, 0), 0)
107 #define LNET_NET_ANY LNET_NIDNET(LNET_NID_ANY)
109 static inline int nid_is_nid4(const struct lnet_nid *nid)
111 return NID_ADDR_BYTES(nid) == 4;
114 /* LOLND may not be defined yet, so we cannot use an inline */
115 #define nid_is_lo0(__nid) \
116 ((__nid)->nid_type == LOLND && \
117 nid_is_nid4(__nid) && \
118 (__nid)->nid_num == 0 && \
119 (__nid)->nid_addr[0] == 0)
121 static inline __u32 LNET_NID_NET(const struct lnet_nid *nid)
123 return LNET_MKNET(nid->nid_type, __be16_to_cpu(nid->nid_num));
126 static inline void lnet_nid4_to_nid(lnet_nid_t nid4, struct lnet_nid *nid)
128 if (nid4 == LNET_NID_ANY) {
129 /* equal to setting to LNET_ANY_NID */
130 memset(nid, 0xff, sizeof(*nid));
135 nid->nid_type = LNET_NETTYP(LNET_NIDNET(nid4));
136 nid->nid_num = __cpu_to_be16(LNET_NETNUM(LNET_NIDNET(nid4)));
137 nid->nid_addr[0] = __cpu_to_be32(LNET_NIDADDR(nid4));
138 nid->nid_addr[1] = nid->nid_addr[2] = nid->nid_addr[3] = 0;
141 static inline lnet_nid_t lnet_nid_to_nid4(const struct lnet_nid *nid)
143 if (LNET_NID_IS_ANY(nid))
146 return LNET_MKNID(LNET_NID_NET(nid), __be32_to_cpu(nid->nid_addr[0]));
149 static inline int nid_same(const struct lnet_nid *n1,
150 const struct lnet_nid *n2)
152 return n1->nid_size == n2->nid_size &&
153 n1->nid_type == n2->nid_type &&
154 n1->nid_num == n2->nid_num &&
155 n1->nid_addr[0] == n2->nid_addr[0] &&
156 n1->nid_addr[1] == n2->nid_addr[1] &&
157 n1->nid_addr[2] == n2->nid_addr[2] &&
158 n1->nid_addr[3] == n2->nid_addr[3];
161 struct lnet_counters_health {
163 __u32 lch_resend_count;
164 __u32 lch_response_timeout_count;
165 __u32 lch_local_interrupt_count;
166 __u32 lch_local_dropped_count;
167 __u32 lch_local_aborted_count;
168 __u32 lch_local_no_route_count;
169 __u32 lch_local_timeout_count;
170 __u32 lch_local_error_count;
171 __u32 lch_remote_dropped_count;
172 __u32 lch_remote_error_count;
173 __u32 lch_remote_timeout_count;
174 __u32 lch_network_timeout_count;
177 struct lnet_counters {
178 struct lnet_counters_common lct_common;
179 struct lnet_counters_health lct_health;
183 * This is a hard-coded limit on the number of interfaces supported by
184 * the interface bonding implemented by the ksocknal LND. It must be
185 * defined here because it is used in LNet data structures that are
186 * common to all LNDs.
188 #define LNET_INTERFACES_NUM 16
190 /* The minimum number of interfaces per node supported by LNet. */
191 #define LNET_INTERFACES_MIN 16
192 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
193 #define LNET_INTERFACES_MAX_DEFAULT 200
196 * Objects maintained by the LNet are accessed through handles. Handle types
197 * have names of the form lnet_handle_xx, where xx is one of the two letter
198 * object type codes ('md' for memory descriptor, and
199 * 'me' for match entry). Each type of object is given a unique handle type
200 * to enhance type checking.
202 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
204 struct lnet_handle_md {
209 * Invalidate md handle \a h.
211 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
213 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
217 * Check whether handler \a h is invalid.
219 * \return 1 if handle is invalid, 0 if valid.
221 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
223 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
229 struct lnet_process_id {
237 /** \addtogroup lnet_me
241 * Specifies whether the match entry or memory descriptor should be unlinked
242 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
250 * Values of the type enum lnet_ins_pos are used to control where a new match
251 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
252 * entry before the current entry or before the head of the list. The value
253 * LNET_INS_AFTER is used to insert the new entry after the current entry
254 * or after the last item in the list.
257 /** insert ME before current position or head of the list */
259 /** insert ME after current position or tail of the list */
261 /** attach ME at tail of local CPU partition ME list */
267 /** \addtogroup lnet_md
271 * Event queue handler function type.
273 * The EQ handler runs for each event that is deposited into the EQ. The
274 * handler is supplied with a pointer to the event that triggered the
275 * handler invocation.
277 * The handler must not block, must be reentrant, and must not call any LNet
278 * API functions. It should return as quickly as possible.
281 typedef void (*lnet_handler_t)(struct lnet_event *event);
284 * Defines the visible parts of a memory descriptor. Values of this type
285 * are used to initialize memory descriptors.
289 * Specify the memory region associated with the memory descriptor.
290 * If the options field has:
291 * - LNET_MD_KIOV bit set: The start field points to the starting
292 * address of an array of struct bio_vec and the length field specifies
293 * the number of entries in the array. The length can't be bigger
294 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
295 * fragments that are not necessarily mapped in virtal memory.
296 * - Otherwise: The memory region is contiguous. The start field
297 * specifies the starting address for the memory region and the
298 * length field specifies its length.
300 * When the memory region is fragmented, all fragments but the first
301 * one must start on page boundary, and all but the last must end on
307 * Specifies the maximum number of operations that can be performed
308 * on the memory descriptor. An operation is any action that could
309 * possibly generate an event. In the usual case, the threshold value
310 * is decremented for each operation on the MD. When the threshold
311 * drops to zero, the MD becomes inactive and does not respond to
312 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
313 * there is no bound on the number of operations that may be applied
318 * Specifies the largest incoming request that the memory descriptor
319 * should respond to. When the unused portion of a MD (length -
320 * local offset) falls below this value, the MD becomes inactive and
321 * does not respond to further operations. This value is only used
322 * if the LNET_MD_MAX_SIZE option is set.
326 * Specifies the behavior of the memory descriptor. A bitwise OR
327 * of the following values can be used:
328 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
329 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
330 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
331 * region is provided by the incoming request. By default, the
332 * offset is maintained locally. When maintained locally, the
333 * offset is incremented by the length of the request so that
334 * the next operation (PUT or GET) will access the next part of
335 * the memory region. Note that only one offset variable exists
336 * per memory descriptor. If both PUT and GET operations are
337 * performed on a memory descriptor, the offset is updated each time.
338 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
339 * be reduced to match the memory available in the region (determined
340 * by subtracting the offset from the length of the memory region).
341 * By default, if the length in the incoming operation is greater
342 * than the amount of memory available, the operation is rejected.
343 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
344 * incoming PUT operations, even if requested. By default,
345 * acknowledgments are sent for PUT operations that request an
346 * acknowledgment. Acknowledgments are never sent for GET operations.
347 * The data sent in the REPLY serves as an implicit acknowledgment.
348 * - LNET_MD_KIOV: The start and length fields specify an array of
350 * - LNET_MD_MAX_SIZE: The max_size field is valid.
351 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
352 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
353 * regardless of the value of the lnet_response_tracking param.
354 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
355 * regardless of the value of the lnet_response_tracking param.
358 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
360 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
361 * region (i.e. sum of all fragment lengths) must not be less than
364 unsigned int options;
366 * A user-specified value that is associated with the memory
367 * descriptor. The value does not need to be a pointer, but must fit
368 * in the space used by a pointer. This value is recorded in events
369 * associated with operations on this MD.
373 * The event handler used to log the operations performed on
374 * the memory region. If this argument is NULL operations
375 * performed on this memory descriptor are not logged.
377 lnet_handler_t handler;
379 * The bulk MD handle which was registered to describe the buffers
380 * either to be used to transfer data to the peer or receive data
381 * from the peer. This allows LNet to properly determine the NUMA
382 * node on which the memory was allocated and use that to select the
383 * nearest local network interface. This value is only used
384 * if the LNET_MD_BULK_HANDLE option is set.
386 struct lnet_handle_md bulk_handle;
389 /* Max Transfer Unit (minimum supported everywhere).
390 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
391 * these limits are system wide and not interface-local. */
392 #define LNET_MTU_BITS 20
393 #define LNET_MTU (1 << LNET_MTU_BITS)
396 * Options for the MD structure. See struct lnet_md::options.
398 #define LNET_MD_OP_PUT (1 << 0)
399 /** See struct lnet_md::options. */
400 #define LNET_MD_OP_GET (1 << 1)
401 /** See struct lnet_md::options. */
402 #define LNET_MD_MANAGE_REMOTE (1 << 2)
403 /* unused (1 << 3) */
404 /** See struct lnet_md::options. */
405 #define LNET_MD_TRUNCATE (1 << 4)
406 /** See struct lnet_md::options. */
407 #define LNET_MD_ACK_DISABLE (1 << 5)
408 /** See struct lnet_md::options. */
409 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
410 /** See struct lnet_md::options. */
411 #define LNET_MD_MAX_SIZE (1 << 7)
412 /** See struct lnet_md::options. */
413 #define LNET_MD_KIOV (1 << 8)
414 /** See struct lnet_md::options. */
415 #define LNET_MD_BULK_HANDLE (1 << 9)
416 /** See struct lnet_md::options. */
417 #define LNET_MD_TRACK_RESPONSE (1 << 10)
418 /** See struct lnet_md::options. */
419 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
421 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
422 #define LNET_MD_THRESH_INF (-1)
426 /** \addtogroup lnet_eq
430 * Six types of events can be logged in an event queue.
432 enum lnet_event_kind {
433 /** An incoming GET operation has completed on the MD. */
436 * An incoming PUT operation has completed on the MD. The
437 * underlying layers will not alter the memory (on behalf of this
438 * operation) once this event has been logged.
442 * A REPLY operation has completed. This event is logged after the
443 * data (if any) from the REPLY has been written into the MD.
446 /** An acknowledgment has been received. */
449 * An outgoing send (PUT or GET) operation has completed. This event
450 * is logged after the entire buffer has been sent and it is safe for
451 * the caller to reuse the buffer.
454 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
455 * happen even when the message has not yet been put out on wire.
456 * - It's unsafe to assume that in an outgoing GET operation
457 * the LNET_EVENT_SEND event would happen before the
458 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
459 * LNET_EVENT_ACK events in an outgoing PUT operation.
463 * A MD has been unlinked. Note that LNetMDUnlink() does not
464 * necessarily trigger an LNET_EVENT_UNLINK event.
470 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
473 * Information about an event on a MD.
476 /** The identifier (nid, pid) of the target. */
477 struct lnet_process_id target;
478 /** The identifier (nid, pid) of the initiator. */
479 struct lnet_process_id initiator;
480 /** The source NID on the initiator. */
481 struct lnet_process_id source;
483 * The NID of the immediate sender. If the request has been forwarded
484 * by routers, this is the NID of the last hop; otherwise it's the
485 * same as the source.
488 /** Indicates the type of the event. */
489 enum lnet_event_kind type;
490 /** The portal table index specified in the request */
491 unsigned int pt_index;
492 /** A copy of the match bits specified in the request. */
494 /** The length (in bytes) specified in the request. */
495 unsigned int rlength;
497 * The length (in bytes) of the data that was manipulated by the
498 * operation. For truncated operations, the manipulated length will be
499 * the number of bytes specified by the MD (possibly with an offset,
500 * see struct lnet_md). For all other operations, the manipulated length
501 * will be the length of the requested operation, i.e. rlength.
503 unsigned int mlength;
505 * The handle to the MD associated with the event. The handle may be
506 * invalid if the MD has been unlinked.
508 struct lnet_handle_md md_handle;
510 * A snapshot of relevant state of the MD immediately after the event
511 * has been processed.
515 unsigned int md_options;
517 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
522 * The message type, to ensure a handler for LNET_EVENT_SEND can
523 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
527 * Indicates the completion status of the operation. It's 0 for
528 * successful operations, otherwise it's an error code.
532 * Indicates whether the MD has been unlinked. Note that:
533 * - An event with unlinked set is the last event on the MD.
534 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
539 * The displacement (in bytes) into the memory region that the
540 * operation used. The offset can be determined by the operation for
541 * a remote managed MD or by the local MD.
542 * \see struct lnet_md::options
546 * The sequence number for this event. Sequence numbers are unique
549 volatile unsigned long sequence;
552 /** \addtogroup lnet_data
556 * Specify whether an acknowledgment should be sent by target when the PUT
557 * operation completes (i.e., when the data has been written to a MD of the
560 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
561 * by which acknowledgments can be disabled for a MD.
564 /** Request an acknowledgment */
566 /** Request that no acknowledgment should be generated. */
571 * UDSP action types. There are two available actions:
572 * 1. PRIORITY - set priority of matching LNet constructs
573 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
575 enum lnet_udsp_action_type {
576 EN_LNET_UDSP_ACTION_NONE = 0,
577 /** assign a priority to matching constructs */
578 EN_LNET_UDSP_ACTION_PRIORITY = 1,
579 /** assign a preferred list of NIDs to matching constructs */
580 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,