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/types.h>
36 #include <linux/string.h>
37 #include <asm/byteorder.h>
45 #include <linux/lnet/lnet-idl.h>
47 /** \addtogroup lnet_addr
50 #define LNET_VERSION "0.7.0"
52 /** Portal reserved for LNet's own use.
53 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
55 #define LNET_RESERVED_PORTAL 0
57 /** wildcard NID that matches any end-point address */
58 #define LNET_NID_ANY (~(lnet_nid_t) 0)
59 /** wildcard PID that matches any lnet_pid_t */
60 #define LNET_PID_ANY (~(lnet_pid_t) 0)
62 static inline int LNET_NID_IS_ANY(const struct lnet_nid *nid)
64 /* A NULL pointer can be used to mean "ANY" */
65 return !nid || nid->nid_type == 0xFF;
68 #define LNET_ANY_NID ((struct lnet_nid) \
69 {0xFF, 0xFF, ~0, {~0, ~0, ~0, ~0} })
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 /* how an LNET NID encodes net:address */
76 /** extract the address part of an lnet_nid_t */
78 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
80 return nid & 0xffffffff;
83 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
85 return (nid >> 32) & 0xffffffff;
88 static inline __u32 LNET_NETNUM(__u32 net)
93 static inline __u32 LNET_NETTYP(__u32 net)
95 return (net >> 16) & 0xff;
98 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
100 return (type << 16) | num;
103 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
105 return (((__u64)net) << 32) | addr;
108 /** The lolnd NID (i.e. myself) */
109 #define LNET_NID_LO_0 LNET_MKNID(LNET_MKNET(LOLND, 0), 0)
111 #define LNET_NET_ANY LNET_NIDNET(LNET_NID_ANY)
113 static inline bool nid_is_nid4(const struct lnet_nid *nid)
115 return NID_ADDR_BYTES(nid) == 4;
118 /* check for address set */
119 static inline bool nid_addr_is_set(const struct lnet_nid *nid)
123 for (i = 0; i < NID_ADDR_BYTES(nid); i++)
124 if (nid->nid_addr[i])
130 /* LOLND may not be defined yet, so we cannot use an inline */
131 #define nid_is_lo0(__nid) \
132 ((__nid)->nid_type == LOLND && \
133 nid_is_nid4(__nid) && \
134 (__nid)->nid_num == 0 && \
135 (__nid)->nid_addr[0] == 0)
137 static inline __u32 LNET_NID_NET(const struct lnet_nid *nid)
139 if (LNET_NID_IS_ANY(nid))
142 return LNET_MKNET(nid->nid_type, __be16_to_cpu(nid->nid_num));
145 static inline void lnet_nid4_to_nid(lnet_nid_t nid4, struct lnet_nid *nid)
147 if (nid4 == LNET_NID_ANY) {
153 nid->nid_type = LNET_NETTYP(LNET_NIDNET(nid4));
154 nid->nid_num = __cpu_to_be16(LNET_NETNUM(LNET_NIDNET(nid4)));
155 nid->nid_addr[0] = __cpu_to_be32(LNET_NIDADDR(nid4));
156 nid->nid_addr[1] = nid->nid_addr[2] = nid->nid_addr[3] = 0;
159 static inline lnet_nid_t lnet_nid_to_nid4(const struct lnet_nid *nid)
161 if (LNET_NID_IS_ANY(nid))
164 return LNET_MKNID(LNET_NID_NET(nid), __be32_to_cpu(nid->nid_addr[0]));
167 static inline int nid_same(const struct lnet_nid *n1,
168 const struct lnet_nid *n2)
170 return n1->nid_size == n2->nid_size &&
171 n1->nid_type == n2->nid_type &&
172 n1->nid_num == n2->nid_num &&
173 n1->nid_addr[0] == n2->nid_addr[0] &&
174 n1->nid_addr[1] == n2->nid_addr[1] &&
175 n1->nid_addr[2] == n2->nid_addr[2] &&
176 n1->nid_addr[3] == n2->nid_addr[3];
179 /* This can be used when we need to hash a nid */
180 static inline unsigned long nidhash(const struct lnet_nid *nid)
183 unsigned long hash = 0;
185 hash ^= LNET_NID_NET(nid);
186 for (i = 0; i < 4; i++)
187 hash ^= nid->nid_addr[i];
191 struct lnet_counters_health {
193 __u32 lch_resend_count;
194 __u32 lch_response_timeout_count;
195 __u32 lch_local_interrupt_count;
196 __u32 lch_local_dropped_count;
197 __u32 lch_local_aborted_count;
198 __u32 lch_local_no_route_count;
199 __u32 lch_local_timeout_count;
200 __u32 lch_local_error_count;
201 __u32 lch_remote_dropped_count;
202 __u32 lch_remote_error_count;
203 __u32 lch_remote_timeout_count;
204 __u32 lch_network_timeout_count;
207 struct lnet_counters {
208 struct lnet_counters_common lct_common;
209 struct lnet_counters_health lct_health;
213 * This is a hard-coded limit on the number of interfaces supported by
214 * the interface bonding implemented by the ksocknal LND. It must be
215 * defined here because it is used in LNet data structures that are
216 * common to all LNDs.
218 #define LNET_INTERFACES_NUM 16
220 /* The minimum number of interfaces per node supported by LNet. */
221 #define LNET_INTERFACES_MIN 16
222 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
223 #define LNET_INTERFACES_MAX_DEFAULT 200
226 * Objects maintained by the LNet are accessed through handles. Handle types
227 * have names of the form lnet_handle_xx, where xx is one of the two letter
228 * object type codes ('md' for memory descriptor, and
229 * 'me' for match entry). Each type of object is given a unique handle type
230 * to enhance type checking.
232 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
234 struct lnet_handle_md {
239 * Invalidate md handle \a h.
241 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
243 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
247 * Check whether handler \a h is invalid.
249 * \return 1 if handle is invalid, 0 if valid.
251 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
253 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
259 struct lnet_process_id {
267 * Global process ID - with large addresses
269 struct lnet_processid {
277 lnet_pid4_to_pid(struct lnet_process_id pid4, struct lnet_processid *pid)
280 lnet_nid4_to_nid(pid4.nid, &pid->nid);
283 static inline struct lnet_process_id
284 lnet_pid_to_pid4(struct lnet_processid *pid)
286 struct lnet_process_id ret;
289 ret.nid = lnet_nid_to_nid4(&pid->nid);
295 /** \addtogroup lnet_me
299 * Specifies whether the match entry or memory descriptor should be unlinked
300 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
308 * Values of the type enum lnet_ins_pos are used to control where a new match
309 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
310 * entry before the current entry or before the head of the list. The value
311 * LNET_INS_AFTER is used to insert the new entry after the current entry
312 * or after the last item in the list.
315 /** insert ME before current position or head of the list */
317 /** insert ME after current position or tail of the list */
319 /** attach ME at tail of local CPU partition ME list */
325 /** \addtogroup lnet_md
328 struct lnet_hdr_nid16 {
329 char _bytes[sizeof(struct lnet_hdr)];
330 } __attribute__((packed));
333 * Event queue handler function type.
335 * The EQ handler runs for each event that is deposited into the EQ. The
336 * handler is supplied with a pointer to the event that triggered the
337 * handler invocation.
339 * The handler must not block, must be reentrant, and must not call any LNet
340 * API functions. It should return as quickly as possible.
343 typedef void (*lnet_handler_t)(struct lnet_event *event);
346 * Defines the visible parts of a memory descriptor. Values of this type
347 * are used to initialize memory descriptors.
351 * Specify the memory region associated with the memory descriptor.
352 * If the options field has:
353 * - LNET_MD_KIOV bit set: The start field points to the starting
354 * address of an array of struct bio_vec and the length field specifies
355 * the number of entries in the array. The length can't be bigger
356 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
357 * fragments that are not necessarily mapped in virtal memory.
358 * - Otherwise: The memory region is contiguous. The start field
359 * specifies the starting address for the memory region and the
360 * length field specifies its length.
362 * When the memory region is fragmented, all fragments but the first
363 * one must start on page boundary, and all but the last must end on
369 * Specifies the maximum number of operations that can be performed
370 * on the memory descriptor. An operation is any action that could
371 * possibly generate an event. In the usual case, the threshold value
372 * is decremented for each operation on the MD. When the threshold
373 * drops to zero, the MD becomes inactive and does not respond to
374 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
375 * there is no bound on the number of operations that may be applied
380 * Specifies the largest incoming request that the memory descriptor
381 * should respond to. When the unused portion of a MD (length -
382 * local offset) falls below this value, the MD becomes inactive and
383 * does not respond to further operations. This value is only used
384 * if the LNET_MD_MAX_SIZE option is set.
388 * Specifies the behavior of the memory descriptor. A bitwise OR
389 * of the following values can be used:
390 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
391 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
392 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
393 * region is provided by the incoming request. By default, the
394 * offset is maintained locally. When maintained locally, the
395 * offset is incremented by the length of the request so that
396 * the next operation (PUT or GET) will access the next part of
397 * the memory region. Note that only one offset variable exists
398 * per memory descriptor. If both PUT and GET operations are
399 * performed on a memory descriptor, the offset is updated each time.
400 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
401 * be reduced to match the memory available in the region (determined
402 * by subtracting the offset from the length of the memory region).
403 * By default, if the length in the incoming operation is greater
404 * than the amount of memory available, the operation is rejected.
405 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
406 * incoming PUT operations, even if requested. By default,
407 * acknowledgments are sent for PUT operations that request an
408 * acknowledgment. Acknowledgments are never sent for GET operations.
409 * The data sent in the REPLY serves as an implicit acknowledgment.
410 * - LNET_MD_KIOV: The start and length fields specify an array of
412 * - LNET_MD_MAX_SIZE: The max_size field is valid.
413 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
414 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
415 * regardless of the value of the lnet_response_tracking param.
416 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
417 * regardless of the value of the lnet_response_tracking param.
418 * - LNET_MD_GNILND: Disable warning about exceeding LNET_MAX_IOV.
421 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
423 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
424 * region (i.e. sum of all fragment lengths) must not be less than
427 unsigned int options;
429 * A user-specified value that is associated with the memory
430 * descriptor. The value does not need to be a pointer, but must fit
431 * in the space used by a pointer. This value is recorded in events
432 * associated with operations on this MD.
436 * The event handler used to log the operations performed on
437 * the memory region. If this argument is NULL operations
438 * performed on this memory descriptor are not logged.
440 lnet_handler_t handler;
442 * The bulk MD handle which was registered to describe the buffers
443 * either to be used to transfer data to the peer or receive data
444 * from the peer. This allows LNet to properly determine the NUMA
445 * node on which the memory was allocated and use that to select the
446 * nearest local network interface. This value is only used
447 * if the LNET_MD_BULK_HANDLE option is set.
449 struct lnet_handle_md bulk_handle;
452 /* Max Transfer Unit (minimum supported everywhere).
453 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
454 * these limits are system wide and not interface-local. */
455 #define LNET_MTU_BITS 20
456 #define LNET_MTU (1 << LNET_MTU_BITS)
459 * Options for the MD structure. See struct lnet_md::options.
461 #define LNET_MD_OP_PUT (1 << 0)
462 /** See struct lnet_md::options. */
463 #define LNET_MD_OP_GET (1 << 1)
464 /** See struct lnet_md::options. */
465 #define LNET_MD_MANAGE_REMOTE (1 << 2)
466 /* unused (1 << 3) */
467 /** See struct lnet_md::options. */
468 #define LNET_MD_TRUNCATE (1 << 4)
469 /** See struct lnet_md::options. */
470 #define LNET_MD_ACK_DISABLE (1 << 5)
471 /** See struct lnet_md::options. */
472 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
473 /** See struct lnet_md::options. */
474 #define LNET_MD_MAX_SIZE (1 << 7)
475 /** See struct lnet_md::options. */
476 #define LNET_MD_KIOV (1 << 8)
477 /** See struct lnet_md::options. */
478 #define LNET_MD_BULK_HANDLE (1 << 9)
479 /** See struct lnet_md::options. */
480 #define LNET_MD_TRACK_RESPONSE (1 << 10)
481 /** See struct lnet_md::options. */
482 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
483 /** See struct lnet_md::options. */
484 #define LNET_MD_GNILND (1 << 12)
485 /** Special page mapping handling */
486 #define LNET_MD_GPU_ADDR (1 << 13)
488 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
489 #define LNET_MD_THRESH_INF (-1)
493 /** \addtogroup lnet_eq
497 * Six types of events can be logged in an event queue.
499 enum lnet_event_kind {
500 /** An incoming GET operation has completed on the MD. */
503 * An incoming PUT operation has completed on the MD. The
504 * underlying layers will not alter the memory (on behalf of this
505 * operation) once this event has been logged.
509 * A REPLY operation has completed. This event is logged after the
510 * data (if any) from the REPLY has been written into the MD.
513 /** An acknowledgment has been received. */
516 * An outgoing send (PUT or GET) operation has completed. This event
517 * is logged after the entire buffer has been sent and it is safe for
518 * the caller to reuse the buffer.
521 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
522 * happen even when the message has not yet been put out on wire.
523 * - It's unsafe to assume that in an outgoing GET operation
524 * the LNET_EVENT_SEND event would happen before the
525 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
526 * LNET_EVENT_ACK events in an outgoing PUT operation.
530 * A MD has been unlinked. Note that LNetMDUnlink() does not
531 * necessarily trigger an LNET_EVENT_UNLINK event.
537 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
540 * Information about an event on a MD.
543 /** The identifier (nid, pid) of the target. */
544 struct lnet_processid target;
545 /** The identifier (nid, pid) of the initiator. */
546 struct lnet_processid initiator;
547 /** The source NID on the initiator. */
548 struct lnet_processid source;
550 * The NID of the immediate sender. If the request has been forwarded
551 * by routers, this is the NID of the last hop; otherwise it's the
552 * same as the source.
554 struct lnet_nid sender;
555 /** Indicates the type of the event. */
556 enum lnet_event_kind type;
557 /** The portal table index specified in the request */
558 unsigned int pt_index;
559 /** A copy of the match bits specified in the request. */
561 /** The length (in bytes) specified in the request. */
562 unsigned int rlength;
564 * The length (in bytes) of the data that was manipulated by the
565 * operation. For truncated operations, the manipulated length will be
566 * the number of bytes specified by the MD (possibly with an offset,
567 * see struct lnet_md). For all other operations, the manipulated length
568 * will be the length of the requested operation, i.e. rlength.
570 unsigned int mlength;
572 * The handle to the MD associated with the event. The handle may be
573 * invalid if the MD has been unlinked.
575 struct lnet_handle_md md_handle;
577 * A snapshot of relevant state of the MD immediately after the event
578 * has been processed.
582 unsigned int md_options;
584 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
589 * The message type, to ensure a handler for LNET_EVENT_SEND can
590 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
594 * Indicates the completion status of the operation. It's 0 for
595 * successful operations, otherwise it's an error code.
599 * Indicates whether the MD has been unlinked. Note that:
600 * - An event with unlinked set is the last event on the MD.
601 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
606 * The displacement (in bytes) into the memory region that the
607 * operation used. The offset can be determined by the operation for
608 * a remote managed MD or by the local MD.
609 * \see struct lnet_md::options
613 * The sequence number for this event. Sequence numbers are unique
616 volatile unsigned long sequence;
619 /** \addtogroup lnet_data
623 * Specify whether an acknowledgment should be sent by target when the PUT
624 * operation completes (i.e., when the data has been written to a MD of the
627 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
628 * by which acknowledgments can be disabled for a MD.
631 /** Request an acknowledgment */
633 /** Request that no acknowledgment should be generated. */
638 * UDSP action types. There are two available actions:
639 * 1. PRIORITY - set priority of matching LNet constructs
640 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
642 enum lnet_udsp_action_type {
643 EN_LNET_UDSP_ACTION_NONE = 0,
644 /** assign a priority to matching constructs */
645 EN_LNET_UDSP_ACTION_PRIORITY = 1,
646 /** assign a preferred list of NIDs to matching constructs */
647 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,