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)
121 __u8 *addr = (__u8 *)(&nid->nid_addr[0]);
124 for (i = 0; i < NID_ADDR_BYTES(nid); i++)
131 /* LOLND may not be defined yet, so we cannot use an inline */
132 #define nid_is_lo0(__nid) \
133 ((__nid)->nid_type == LOLND && \
134 nid_is_nid4(__nid) && \
135 (__nid)->nid_num == 0 && \
136 (__nid)->nid_addr[0] == 0)
138 static inline __u32 LNET_NID_NET(const struct lnet_nid *nid)
140 if (LNET_NID_IS_ANY(nid))
143 return LNET_MKNET(nid->nid_type, __be16_to_cpu(nid->nid_num));
146 static inline void lnet_nid4_to_nid(lnet_nid_t nid4, struct lnet_nid *nid)
148 if (nid4 == LNET_NID_ANY) {
154 nid->nid_type = LNET_NETTYP(LNET_NIDNET(nid4));
155 nid->nid_num = __cpu_to_be16(LNET_NETNUM(LNET_NIDNET(nid4)));
156 nid->nid_addr[0] = __cpu_to_be32(LNET_NIDADDR(nid4));
157 nid->nid_addr[1] = nid->nid_addr[2] = nid->nid_addr[3] = 0;
160 static inline lnet_nid_t lnet_nid_to_nid4(const struct lnet_nid *nid)
162 if (LNET_NID_IS_ANY(nid))
165 return LNET_MKNID(LNET_NID_NET(nid), __be32_to_cpu(nid->nid_addr[0]));
168 static inline int nid_same(const struct lnet_nid *n1,
169 const struct lnet_nid *n2)
171 return n1->nid_size == n2->nid_size &&
172 n1->nid_type == n2->nid_type &&
173 n1->nid_num == n2->nid_num &&
174 n1->nid_addr[0] == n2->nid_addr[0] &&
175 n1->nid_addr[1] == n2->nid_addr[1] &&
176 n1->nid_addr[2] == n2->nid_addr[2] &&
177 n1->nid_addr[3] == n2->nid_addr[3];
180 /* This can be used when we need to hash a nid */
181 static inline unsigned long nidhash(const struct lnet_nid *nid)
184 unsigned long hash = 0;
186 hash ^= LNET_NID_NET(nid);
187 for (i = 0; i < 4; i++)
188 hash ^= nid->nid_addr[i];
192 struct lnet_counters_health {
194 __u32 lch_resend_count;
195 __u32 lch_response_timeout_count;
196 __u32 lch_local_interrupt_count;
197 __u32 lch_local_dropped_count;
198 __u32 lch_local_aborted_count;
199 __u32 lch_local_no_route_count;
200 __u32 lch_local_timeout_count;
201 __u32 lch_local_error_count;
202 __u32 lch_remote_dropped_count;
203 __u32 lch_remote_error_count;
204 __u32 lch_remote_timeout_count;
205 __u32 lch_network_timeout_count;
208 struct lnet_counters {
209 struct lnet_counters_common lct_common;
210 struct lnet_counters_health lct_health;
214 * This is a hard-coded limit on the number of interfaces supported by
215 * the interface bonding implemented by the ksocknal LND. It must be
216 * defined here because it is used in LNet data structures that are
217 * common to all LNDs.
219 #define LNET_INTERFACES_NUM 16
221 /* The minimum number of interfaces per node supported by LNet. */
222 #define LNET_INTERFACES_MIN 16
223 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
224 #define LNET_INTERFACES_MAX_DEFAULT 200
227 * Objects maintained by the LNet are accessed through handles. Handle types
228 * have names of the form lnet_handle_xx, where xx is one of the two letter
229 * object type codes ('md' for memory descriptor, and
230 * 'me' for match entry). Each type of object is given a unique handle type
231 * to enhance type checking.
233 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
235 struct lnet_handle_md {
240 * Invalidate md handle \a h.
242 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
244 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
248 * Check whether handler \a h is invalid.
250 * \return 1 if handle is invalid, 0 if valid.
252 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
254 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
260 struct lnet_process_id {
268 * Global process ID - with large addresses
270 struct lnet_processid {
278 lnet_pid4_to_pid(struct lnet_process_id pid4, struct lnet_processid *pid)
281 lnet_nid4_to_nid(pid4.nid, &pid->nid);
284 static inline struct lnet_process_id
285 lnet_pid_to_pid4(struct lnet_processid *pid)
287 struct lnet_process_id ret;
290 ret.nid = lnet_nid_to_nid4(&pid->nid);
296 /** \addtogroup lnet_me
300 * Specifies whether the match entry or memory descriptor should be unlinked
301 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
309 * Values of the type enum lnet_ins_pos are used to control where a new match
310 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
311 * entry before the current entry or before the head of the list. The value
312 * LNET_INS_AFTER is used to insert the new entry after the current entry
313 * or after the last item in the list.
316 /** insert ME before current position or head of the list */
318 /** insert ME after current position or tail of the list */
320 /** attach ME at tail of local CPU partition ME list */
326 /** \addtogroup lnet_md
329 struct lnet_hdr_nid16 {
330 char _bytes[sizeof(struct lnet_hdr)];
331 } __attribute__((packed));
334 * Event queue handler function type.
336 * The EQ handler runs for each event that is deposited into the EQ. The
337 * handler is supplied with a pointer to the event that triggered the
338 * handler invocation.
340 * The handler must not block, must be reentrant, and must not call any LNet
341 * API functions. It should return as quickly as possible.
344 typedef void (*lnet_handler_t)(struct lnet_event *event);
347 * Defines the visible parts of a memory descriptor. Values of this type
348 * are used to initialize memory descriptors.
352 * Specify the memory region associated with the memory descriptor.
353 * If the options field has:
354 * - LNET_MD_KIOV bit set: The start field points to the starting
355 * address of an array of struct bio_vec and the length field specifies
356 * the number of entries in the array. The length can't be bigger
357 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
358 * fragments that are not necessarily mapped in virtal memory.
359 * - Otherwise: The memory region is contiguous. The start field
360 * specifies the starting address for the memory region and the
361 * length field specifies its length.
363 * When the memory region is fragmented, all fragments but the first
364 * one must start on page boundary, and all but the last must end on
370 * Specifies the maximum number of operations that can be performed
371 * on the memory descriptor. An operation is any action that could
372 * possibly generate an event. In the usual case, the threshold value
373 * is decremented for each operation on the MD. When the threshold
374 * drops to zero, the MD becomes inactive and does not respond to
375 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
376 * there is no bound on the number of operations that may be applied
381 * Specifies the largest incoming request that the memory descriptor
382 * should respond to. When the unused portion of a MD (length -
383 * local offset) falls below this value, the MD becomes inactive and
384 * does not respond to further operations. This value is only used
385 * if the LNET_MD_MAX_SIZE option is set.
389 * Specifies the behavior of the memory descriptor. A bitwise OR
390 * of the following values can be used:
391 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
392 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
393 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
394 * region is provided by the incoming request. By default, the
395 * offset is maintained locally. When maintained locally, the
396 * offset is incremented by the length of the request so that
397 * the next operation (PUT or GET) will access the next part of
398 * the memory region. Note that only one offset variable exists
399 * per memory descriptor. If both PUT and GET operations are
400 * performed on a memory descriptor, the offset is updated each time.
401 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
402 * be reduced to match the memory available in the region (determined
403 * by subtracting the offset from the length of the memory region).
404 * By default, if the length in the incoming operation is greater
405 * than the amount of memory available, the operation is rejected.
406 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
407 * incoming PUT operations, even if requested. By default,
408 * acknowledgments are sent for PUT operations that request an
409 * acknowledgment. Acknowledgments are never sent for GET operations.
410 * The data sent in the REPLY serves as an implicit acknowledgment.
411 * - LNET_MD_KIOV: The start and length fields specify an array of
413 * - LNET_MD_MAX_SIZE: The max_size field is valid.
414 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
415 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
416 * regardless of the value of the lnet_response_tracking param.
417 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
418 * regardless of the value of the lnet_response_tracking param.
419 * - LNET_MD_GNILND: Disable warning about exceeding LNET_MAX_IOV.
422 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
424 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
425 * region (i.e. sum of all fragment lengths) must not be less than
428 unsigned int options;
430 * A user-specified value that is associated with the memory
431 * descriptor. The value does not need to be a pointer, but must fit
432 * in the space used by a pointer. This value is recorded in events
433 * associated with operations on this MD.
437 * The event handler used to log the operations performed on
438 * the memory region. If this argument is NULL operations
439 * performed on this memory descriptor are not logged.
441 lnet_handler_t handler;
443 * The bulk MD handle which was registered to describe the buffers
444 * either to be used to transfer data to the peer or receive data
445 * from the peer. This allows LNet to properly determine the NUMA
446 * node on which the memory was allocated and use that to select the
447 * nearest local network interface. This value is only used
448 * if the LNET_MD_BULK_HANDLE option is set.
450 struct lnet_handle_md bulk_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)
460 * Options for the MD structure. See struct lnet_md::options.
462 #define LNET_MD_OP_PUT (1 << 0)
463 /** See struct lnet_md::options. */
464 #define LNET_MD_OP_GET (1 << 1)
465 /** See struct lnet_md::options. */
466 #define LNET_MD_MANAGE_REMOTE (1 << 2)
467 /* unused (1 << 3) */
468 /** See struct lnet_md::options. */
469 #define LNET_MD_TRUNCATE (1 << 4)
470 /** See struct lnet_md::options. */
471 #define LNET_MD_ACK_DISABLE (1 << 5)
472 /** See struct lnet_md::options. */
473 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
474 /** See struct lnet_md::options. */
475 #define LNET_MD_MAX_SIZE (1 << 7)
476 /** See struct lnet_md::options. */
477 #define LNET_MD_KIOV (1 << 8)
478 /** See struct lnet_md::options. */
479 #define LNET_MD_BULK_HANDLE (1 << 9)
480 /** See struct lnet_md::options. */
481 #define LNET_MD_TRACK_RESPONSE (1 << 10)
482 /** See struct lnet_md::options. */
483 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
484 /** See struct lnet_md::options. */
485 #define LNET_MD_GNILND (1 << 12)
486 /** Special page mapping handling */
487 #define LNET_MD_GPU_ADDR (1 << 13)
489 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
490 #define LNET_MD_THRESH_INF (-1)
494 /** \addtogroup lnet_eq
498 * Six types of events can be logged in an event queue.
500 enum lnet_event_kind {
501 /** An incoming GET operation has completed on the MD. */
504 * An incoming PUT operation has completed on the MD. The
505 * underlying layers will not alter the memory (on behalf of this
506 * operation) once this event has been logged.
510 * A REPLY operation has completed. This event is logged after the
511 * data (if any) from the REPLY has been written into the MD.
514 /** An acknowledgment has been received. */
517 * An outgoing send (PUT or GET) operation has completed. This event
518 * is logged after the entire buffer has been sent and it is safe for
519 * the caller to reuse the buffer.
522 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
523 * happen even when the message has not yet been put out on wire.
524 * - It's unsafe to assume that in an outgoing GET operation
525 * the LNET_EVENT_SEND event would happen before the
526 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
527 * LNET_EVENT_ACK events in an outgoing PUT operation.
531 * A MD has been unlinked. Note that LNetMDUnlink() does not
532 * necessarily trigger an LNET_EVENT_UNLINK event.
538 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
541 * Information about an event on a MD.
544 /** The identifier (nid, pid) of the target. */
545 struct lnet_processid target;
546 /** The identifier (nid, pid) of the initiator. */
547 struct lnet_processid initiator;
548 /** The source NID on the initiator. */
549 struct lnet_processid source;
551 * The NID of the immediate sender. If the request has been forwarded
552 * by routers, this is the NID of the last hop; otherwise it's the
553 * same as the source.
555 struct lnet_nid sender;
556 /** Indicates the type of the event. */
557 enum lnet_event_kind type;
558 /** The portal table index specified in the request */
559 unsigned int pt_index;
560 /** A copy of the match bits specified in the request. */
562 /** The length (in bytes) specified in the request. */
563 unsigned int rlength;
565 * The length (in bytes) of the data that was manipulated by the
566 * operation. For truncated operations, the manipulated length will be
567 * the number of bytes specified by the MD (possibly with an offset,
568 * see struct lnet_md). For all other operations, the manipulated length
569 * will be the length of the requested operation, i.e. rlength.
571 unsigned int mlength;
573 * The handle to the MD associated with the event. The handle may be
574 * invalid if the MD has been unlinked.
576 struct lnet_handle_md md_handle;
578 * A snapshot of relevant state of the MD immediately after the event
579 * has been processed.
583 unsigned int md_options;
585 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
590 * The message type, to ensure a handler for LNET_EVENT_SEND can
591 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
595 * Indicates the completion status of the operation. It's 0 for
596 * successful operations, otherwise it's an error code.
600 * Indicates whether the MD has been unlinked. Note that:
601 * - An event with unlinked set is the last event on the MD.
602 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
607 * The displacement (in bytes) into the memory region that the
608 * operation used. The offset can be determined by the operation for
609 * a remote managed MD or by the local MD.
610 * \see struct lnet_md::options
614 * The sequence number for this event. Sequence numbers are unique
617 volatile unsigned long sequence;
620 /** \addtogroup lnet_data
624 * Specify whether an acknowledgment should be sent by target when the PUT
625 * operation completes (i.e., when the data has been written to a MD of the
628 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
629 * by which acknowledgments can be disabled for a MD.
632 /** Request an acknowledgment */
634 /** Request that no acknowledgment should be generated. */
639 * UDSP action types. There are two available actions:
640 * 1. PRIORITY - set priority of matching LNet constructs
641 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
643 enum lnet_udsp_action_type {
644 EN_LNET_UDSP_ACTION_NONE = 0,
645 /** assign a priority to matching constructs */
646 EN_LNET_UDSP_ACTION_PRIORITY = 1,
647 /** assign a preferred list of NIDs to matching constructs */
648 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,