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) -1)
59 /** wildcard PID that matches any lnet_pid_t */
60 #define LNET_PID_ANY ((lnet_pid_t) -1)
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 static inline bool nid_is_ipv4(const struct lnet_nid *nid)
120 return NID_ADDR_BYTES(nid) == 4;
123 static inline bool nid_is_ipv6(const struct lnet_nid *nid)
125 return NID_ADDR_BYTES(nid) == 16;
128 /* check for address set */
129 static inline bool nid_addr_is_set(const struct lnet_nid *nid)
133 for (i = 0; i < NID_ADDR_BYTES(nid); i++)
134 if (nid->nid_addr[i])
140 /* LOLND may not be defined yet, so we cannot use an inline */
141 #define nid_is_lo0(__nid) \
142 ((__nid)->nid_type == LOLND && \
143 nid_is_nid4(__nid) && \
144 (__nid)->nid_num == 0 && \
145 (__nid)->nid_addr[0] == 0)
147 static inline __u32 LNET_NID_NET(const struct lnet_nid *nid)
149 if (LNET_NID_IS_ANY(nid))
152 return LNET_MKNET(nid->nid_type, __be16_to_cpu(nid->nid_num));
155 static inline void lnet_nid4_to_nid(lnet_nid_t nid4, struct lnet_nid *nid)
157 if (nid4 == LNET_NID_ANY) {
158 /* equal to setting to LNET_ANY_NID */
159 memset(nid, 0xff, sizeof(*nid));
164 nid->nid_type = LNET_NETTYP(LNET_NIDNET(nid4));
165 nid->nid_num = __cpu_to_be16(LNET_NETNUM(LNET_NIDNET(nid4)));
166 nid->nid_addr[0] = __cpu_to_be32(LNET_NIDADDR(nid4));
167 nid->nid_addr[1] = nid->nid_addr[2] = nid->nid_addr[3] = 0;
170 static inline lnet_nid_t lnet_nid_to_nid4(const struct lnet_nid *nid)
172 if (LNET_NID_IS_ANY(nid))
175 return LNET_MKNID(LNET_NID_NET(nid), __be32_to_cpu(nid->nid_addr[0]));
178 static inline int nid_same(const struct lnet_nid *n1,
179 const struct lnet_nid *n2)
181 return n1->nid_size == n2->nid_size &&
182 n1->nid_type == n2->nid_type &&
183 n1->nid_num == n2->nid_num &&
184 n1->nid_addr[0] == n2->nid_addr[0] &&
185 n1->nid_addr[1] == n2->nid_addr[1] &&
186 n1->nid_addr[2] == n2->nid_addr[2] &&
187 n1->nid_addr[3] == n2->nid_addr[3];
190 /* This can be used when we need to hash a nid */
191 static inline unsigned long nidhash(const struct lnet_nid *nid)
194 unsigned long hash = 0;
196 hash ^= LNET_NID_NET(nid);
197 for (i = 0; i < 4; i++)
198 hash ^= nid->nid_addr[i];
202 struct lnet_counters_health {
204 __u32 lch_resend_count;
205 __u32 lch_response_timeout_count;
206 __u32 lch_local_interrupt_count;
207 __u32 lch_local_dropped_count;
208 __u32 lch_local_aborted_count;
209 __u32 lch_local_no_route_count;
210 __u32 lch_local_timeout_count;
211 __u32 lch_local_error_count;
212 __u32 lch_remote_dropped_count;
213 __u32 lch_remote_error_count;
214 __u32 lch_remote_timeout_count;
215 __u32 lch_network_timeout_count;
218 struct lnet_counters {
219 struct lnet_counters_common lct_common;
220 struct lnet_counters_health lct_health;
224 * This is a hard-coded limit on the number of interfaces supported by
225 * the interface bonding implemented by the ksocknal LND. It must be
226 * defined here because it is used in LNet data structures that are
227 * common to all LNDs.
229 #define LNET_INTERFACES_NUM 16
231 /* The minimum number of interfaces per node supported by LNet. */
232 #define LNET_INTERFACES_MIN 16
233 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
234 #define LNET_INTERFACES_MAX_DEFAULT 200
237 * Objects maintained by the LNet are accessed through handles. Handle types
238 * have names of the form lnet_handle_xx, where xx is one of the two letter
239 * object type codes ('md' for memory descriptor, and
240 * 'me' for match entry). Each type of object is given a unique handle type
241 * to enhance type checking.
243 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
245 struct lnet_handle_md {
250 * Invalidate md handle \a h.
252 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
254 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
258 * Check whether handler \a h is invalid.
260 * \return 1 if handle is invalid, 0 if valid.
262 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
264 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
270 struct lnet_process_id {
278 * Global process ID - with large addresses
280 struct lnet_processid {
288 lnet_pid4_to_pid(struct lnet_process_id pid4, struct lnet_processid *pid)
291 lnet_nid4_to_nid(pid4.nid, &pid->nid);
294 static inline struct lnet_process_id
295 lnet_pid_to_pid4(struct lnet_processid *pid)
297 struct lnet_process_id ret;
300 ret.nid = lnet_nid_to_nid4(&pid->nid);
306 /** \addtogroup lnet_me
310 * Specifies whether the match entry or memory descriptor should be unlinked
311 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
319 * Values of the type enum lnet_ins_pos are used to control where a new match
320 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
321 * entry before the current entry or before the head of the list. The value
322 * LNET_INS_AFTER is used to insert the new entry after the current entry
323 * or after the last item in the list.
326 /** insert ME before current position or head of the list */
328 /** insert ME after current position or tail of the list */
330 /** attach ME at tail of local CPU partition ME list */
336 /** \addtogroup lnet_md
339 struct lnet_hdr_nid16 {
340 char _bytes[sizeof(struct lnet_hdr)];
341 } __attribute__((packed));
344 * Event queue handler function type.
346 * The EQ handler runs for each event that is deposited into the EQ. The
347 * handler is supplied with a pointer to the event that triggered the
348 * handler invocation.
350 * The handler must not block, must be reentrant, and must not call any LNet
351 * API functions. It should return as quickly as possible.
354 typedef void (*lnet_handler_t)(struct lnet_event *event);
357 * Defines the visible parts of a memory descriptor. Values of this type
358 * are used to initialize memory descriptors.
362 * Specify the memory region associated with the memory descriptor.
363 * If the options field has:
364 * - LNET_MD_KIOV bit set: The start field points to the starting
365 * address of an array of struct bio_vec and the length field specifies
366 * the number of entries in the array. The length can't be bigger
367 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
368 * fragments that are not necessarily mapped in virtal memory.
369 * - Otherwise: The memory region is contiguous. The start field
370 * specifies the starting address for the memory region and the
371 * length field specifies its length.
373 * When the memory region is fragmented, all fragments but the first
374 * one must start on page boundary, and all but the last must end on
380 * Specifies the maximum number of operations that can be performed
381 * on the memory descriptor. An operation is any action that could
382 * possibly generate an event. In the usual case, the threshold value
383 * is decremented for each operation on the MD. When the threshold
384 * drops to zero, the MD becomes inactive and does not respond to
385 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
386 * there is no bound on the number of operations that may be applied
391 * Specifies the largest incoming request that the memory descriptor
392 * should respond to. When the unused portion of a MD (length -
393 * local offset) falls below this value, the MD becomes inactive and
394 * does not respond to further operations. This value is only used
395 * if the LNET_MD_MAX_SIZE option is set.
399 * Specifies the behavior of the memory descriptor. A bitwise OR
400 * of the following values can be used:
401 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
402 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
403 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
404 * region is provided by the incoming request. By default, the
405 * offset is maintained locally. When maintained locally, the
406 * offset is incremented by the length of the request so that
407 * the next operation (PUT or GET) will access the next part of
408 * the memory region. Note that only one offset variable exists
409 * per memory descriptor. If both PUT and GET operations are
410 * performed on a memory descriptor, the offset is updated each time.
411 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
412 * be reduced to match the memory available in the region (determined
413 * by subtracting the offset from the length of the memory region).
414 * By default, if the length in the incoming operation is greater
415 * than the amount of memory available, the operation is rejected.
416 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
417 * incoming PUT operations, even if requested. By default,
418 * acknowledgments are sent for PUT operations that request an
419 * acknowledgment. Acknowledgments are never sent for GET operations.
420 * The data sent in the REPLY serves as an implicit acknowledgment.
421 * - LNET_MD_KIOV: The start and length fields specify an array of
423 * - LNET_MD_MAX_SIZE: The max_size field is valid.
424 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
425 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
426 * regardless of the value of the lnet_response_tracking param.
427 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
428 * regardless of the value of the lnet_response_tracking param.
429 * - LNET_MD_GNILND: Disable warning about exceeding LNET_MAX_IOV.
432 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
434 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
435 * region (i.e. sum of all fragment lengths) must not be less than
438 unsigned int options;
440 * A user-specified value that is associated with the memory
441 * descriptor. The value does not need to be a pointer, but must fit
442 * in the space used by a pointer. This value is recorded in events
443 * associated with operations on this MD.
447 * The event handler used to log the operations performed on
448 * the memory region. If this argument is NULL operations
449 * performed on this memory descriptor are not logged.
451 lnet_handler_t handler;
453 * The bulk MD handle which was registered to describe the buffers
454 * either to be used to transfer data to the peer or receive data
455 * from the peer. This allows LNet to properly determine the NUMA
456 * node on which the memory was allocated and use that to select the
457 * nearest local network interface. This value is only used
458 * if the LNET_MD_BULK_HANDLE option is set.
460 struct lnet_handle_md bulk_handle;
463 /* Max Transfer Unit (minimum supported everywhere).
464 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
465 * these limits are system wide and not interface-local. */
466 #define LNET_MTU_BITS 20
467 #define LNET_MTU (1 << LNET_MTU_BITS)
470 * Options for the MD structure. See struct lnet_md::options.
472 #define LNET_MD_OP_PUT (1 << 0)
473 /** See struct lnet_md::options. */
474 #define LNET_MD_OP_GET (1 << 1)
475 /** See struct lnet_md::options. */
476 #define LNET_MD_MANAGE_REMOTE (1 << 2)
477 /* unused (1 << 3) */
478 /** See struct lnet_md::options. */
479 #define LNET_MD_TRUNCATE (1 << 4)
480 /** See struct lnet_md::options. */
481 #define LNET_MD_ACK_DISABLE (1 << 5)
482 /** See struct lnet_md::options. */
483 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
484 /** See struct lnet_md::options. */
485 #define LNET_MD_MAX_SIZE (1 << 7)
486 /** See struct lnet_md::options. */
487 #define LNET_MD_KIOV (1 << 8)
488 /** See struct lnet_md::options. */
489 #define LNET_MD_BULK_HANDLE (1 << 9)
490 /** See struct lnet_md::options. */
491 #define LNET_MD_TRACK_RESPONSE (1 << 10)
492 /** See struct lnet_md::options. */
493 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
494 /** See struct lnet_md::options. */
495 #define LNET_MD_GNILND (1 << 12)
496 /** Special page mapping handling */
497 #define LNET_MD_GPU_ADDR (1 << 13)
499 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
500 #define LNET_MD_THRESH_INF (-1)
504 /** \addtogroup lnet_eq
508 * Six types of events can be logged in an event queue.
510 enum lnet_event_kind {
511 /** An incoming GET operation has completed on the MD. */
514 * An incoming PUT operation has completed on the MD. The
515 * underlying layers will not alter the memory (on behalf of this
516 * operation) once this event has been logged.
520 * A REPLY operation has completed. This event is logged after the
521 * data (if any) from the REPLY has been written into the MD.
524 /** An acknowledgment has been received. */
527 * An outgoing send (PUT or GET) operation has completed. This event
528 * is logged after the entire buffer has been sent and it is safe for
529 * the caller to reuse the buffer.
532 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
533 * happen even when the message has not yet been put out on wire.
534 * - It's unsafe to assume that in an outgoing GET operation
535 * the LNET_EVENT_SEND event would happen before the
536 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
537 * LNET_EVENT_ACK events in an outgoing PUT operation.
541 * A MD has been unlinked. Note that LNetMDUnlink() does not
542 * necessarily trigger an LNET_EVENT_UNLINK event.
548 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
551 * Information about an event on a MD.
554 /** The identifier (nid, pid) of the target. */
555 struct lnet_processid target;
556 /** The identifier (nid, pid) of the initiator. */
557 struct lnet_processid initiator;
558 /** The source NID on the initiator. */
559 struct lnet_processid source;
561 * The NID of the immediate sender. If the request has been forwarded
562 * by routers, this is the NID of the last hop; otherwise it's the
563 * same as the source.
565 struct lnet_nid sender;
566 /** Indicates the type of the event. */
567 enum lnet_event_kind type;
568 /** The portal table index specified in the request */
569 unsigned int pt_index;
570 /** A copy of the match bits specified in the request. */
572 /** The length (in bytes) specified in the request. */
573 unsigned int rlength;
575 * The length (in bytes) of the data that was manipulated by the
576 * operation. For truncated operations, the manipulated length will be
577 * the number of bytes specified by the MD (possibly with an offset,
578 * see struct lnet_md). For all other operations, the manipulated length
579 * will be the length of the requested operation, i.e. rlength.
581 unsigned int mlength;
583 * The handle to the MD associated with the event. The handle may be
584 * invalid if the MD has been unlinked.
586 struct lnet_handle_md md_handle;
588 * A snapshot of relevant state of the MD immediately after the event
589 * has been processed.
593 unsigned int md_options;
595 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
600 * The message type, to ensure a handler for LNET_EVENT_SEND can
601 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
605 * Indicates the completion status of the operation. It's 0 for
606 * successful operations, otherwise it's an error code.
610 * Indicates whether the MD has been unlinked. Note that:
611 * - An event with unlinked set is the last event on the MD.
612 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
617 * The displacement (in bytes) into the memory region that the
618 * operation used. The offset can be determined by the operation for
619 * a remote managed MD or by the local MD.
620 * \see struct lnet_md::options
624 * The sequence number for this event. Sequence numbers are unique
627 volatile unsigned long sequence;
630 /** \addtogroup lnet_data
634 * Specify whether an acknowledgment should be sent by target when the PUT
635 * operation completes (i.e., when the data has been written to a MD of the
638 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
639 * by which acknowledgments can be disabled for a MD.
642 /** Request an acknowledgment */
644 /** Request that no acknowledgment should be generated. */
649 * UDSP action types. There are two available actions:
650 * 1. PRIORITY - set priority of matching LNet constructs
651 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
653 enum lnet_udsp_action_type {
654 EN_LNET_UDSP_ACTION_NONE = 0,
655 /** assign a priority to matching constructs */
656 EN_LNET_UDSP_ACTION_PRIORITY = 1,
657 /** assign a preferred list of NIDs to matching constructs */
658 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,