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 /* This can be used when we need to hash a nid */
162 static inline unsigned long nidhash(const struct lnet_nid *nid)
165 unsigned long hash = 0;
167 hash ^= LNET_NID_NET(nid);
168 for (i = 0; i < 4; i++)
169 hash ^= nid->nid_addr[i];
173 struct lnet_counters_health {
175 __u32 lch_resend_count;
176 __u32 lch_response_timeout_count;
177 __u32 lch_local_interrupt_count;
178 __u32 lch_local_dropped_count;
179 __u32 lch_local_aborted_count;
180 __u32 lch_local_no_route_count;
181 __u32 lch_local_timeout_count;
182 __u32 lch_local_error_count;
183 __u32 lch_remote_dropped_count;
184 __u32 lch_remote_error_count;
185 __u32 lch_remote_timeout_count;
186 __u32 lch_network_timeout_count;
189 struct lnet_counters {
190 struct lnet_counters_common lct_common;
191 struct lnet_counters_health lct_health;
195 * This is a hard-coded limit on the number of interfaces supported by
196 * the interface bonding implemented by the ksocknal LND. It must be
197 * defined here because it is used in LNet data structures that are
198 * common to all LNDs.
200 #define LNET_INTERFACES_NUM 16
202 /* The minimum number of interfaces per node supported by LNet. */
203 #define LNET_INTERFACES_MIN 16
204 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
205 #define LNET_INTERFACES_MAX_DEFAULT 200
208 * Objects maintained by the LNet are accessed through handles. Handle types
209 * have names of the form lnet_handle_xx, where xx is one of the two letter
210 * object type codes ('md' for memory descriptor, and
211 * 'me' for match entry). Each type of object is given a unique handle type
212 * to enhance type checking.
214 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
216 struct lnet_handle_md {
221 * Invalidate md handle \a h.
223 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
225 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
229 * Check whether handler \a h is invalid.
231 * \return 1 if handle is invalid, 0 if valid.
233 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
235 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
241 struct lnet_process_id {
249 * Global process ID - with large addresses
251 struct lnet_processid {
259 lnet_pid4_to_pid(struct lnet_process_id pid4, struct lnet_processid *pid)
262 lnet_nid4_to_nid(pid4.nid, &pid->nid);
265 static inline struct lnet_process_id
266 lnet_pid_to_pid4(struct lnet_processid *pid)
268 struct lnet_process_id ret;
271 ret.nid = lnet_nid_to_nid4(&pid->nid);
277 /** \addtogroup lnet_me
281 * Specifies whether the match entry or memory descriptor should be unlinked
282 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
290 * Values of the type enum lnet_ins_pos are used to control where a new match
291 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
292 * entry before the current entry or before the head of the list. The value
293 * LNET_INS_AFTER is used to insert the new entry after the current entry
294 * or after the last item in the list.
297 /** insert ME before current position or head of the list */
299 /** insert ME after current position or tail of the list */
301 /** attach ME at tail of local CPU partition ME list */
307 /** \addtogroup lnet_md
310 struct lnet_hdr_nid16 {
311 char _bytes[sizeof(struct lnet_hdr)];
312 } __attribute__((packed));
315 * Event queue handler function type.
317 * The EQ handler runs for each event that is deposited into the EQ. The
318 * handler is supplied with a pointer to the event that triggered the
319 * handler invocation.
321 * The handler must not block, must be reentrant, and must not call any LNet
322 * API functions. It should return as quickly as possible.
325 typedef void (*lnet_handler_t)(struct lnet_event *event);
328 * Defines the visible parts of a memory descriptor. Values of this type
329 * are used to initialize memory descriptors.
333 * Specify the memory region associated with the memory descriptor.
334 * If the options field has:
335 * - LNET_MD_KIOV bit set: The start field points to the starting
336 * address of an array of struct bio_vec and the length field specifies
337 * the number of entries in the array. The length can't be bigger
338 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
339 * fragments that are not necessarily mapped in virtal memory.
340 * - Otherwise: The memory region is contiguous. The start field
341 * specifies the starting address for the memory region and the
342 * length field specifies its length.
344 * When the memory region is fragmented, all fragments but the first
345 * one must start on page boundary, and all but the last must end on
351 * Specifies the maximum number of operations that can be performed
352 * on the memory descriptor. An operation is any action that could
353 * possibly generate an event. In the usual case, the threshold value
354 * is decremented for each operation on the MD. When the threshold
355 * drops to zero, the MD becomes inactive and does not respond to
356 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
357 * there is no bound on the number of operations that may be applied
362 * Specifies the largest incoming request that the memory descriptor
363 * should respond to. When the unused portion of a MD (length -
364 * local offset) falls below this value, the MD becomes inactive and
365 * does not respond to further operations. This value is only used
366 * if the LNET_MD_MAX_SIZE option is set.
370 * Specifies the behavior of the memory descriptor. A bitwise OR
371 * of the following values can be used:
372 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
373 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
374 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
375 * region is provided by the incoming request. By default, the
376 * offset is maintained locally. When maintained locally, the
377 * offset is incremented by the length of the request so that
378 * the next operation (PUT or GET) will access the next part of
379 * the memory region. Note that only one offset variable exists
380 * per memory descriptor. If both PUT and GET operations are
381 * performed on a memory descriptor, the offset is updated each time.
382 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
383 * be reduced to match the memory available in the region (determined
384 * by subtracting the offset from the length of the memory region).
385 * By default, if the length in the incoming operation is greater
386 * than the amount of memory available, the operation is rejected.
387 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
388 * incoming PUT operations, even if requested. By default,
389 * acknowledgments are sent for PUT operations that request an
390 * acknowledgment. Acknowledgments are never sent for GET operations.
391 * The data sent in the REPLY serves as an implicit acknowledgment.
392 * - LNET_MD_KIOV: The start and length fields specify an array of
394 * - LNET_MD_MAX_SIZE: The max_size field is valid.
395 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
396 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
397 * regardless of the value of the lnet_response_tracking param.
398 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
399 * regardless of the value of the lnet_response_tracking param.
400 * - LNET_MD_GNILND: Disable warning about exceeding LNET_MAX_IOV.
403 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
405 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
406 * region (i.e. sum of all fragment lengths) must not be less than
409 unsigned int options;
411 * A user-specified value that is associated with the memory
412 * descriptor. The value does not need to be a pointer, but must fit
413 * in the space used by a pointer. This value is recorded in events
414 * associated with operations on this MD.
418 * The event handler used to log the operations performed on
419 * the memory region. If this argument is NULL operations
420 * performed on this memory descriptor are not logged.
422 lnet_handler_t handler;
424 * The bulk MD handle which was registered to describe the buffers
425 * either to be used to transfer data to the peer or receive data
426 * from the peer. This allows LNet to properly determine the NUMA
427 * node on which the memory was allocated and use that to select the
428 * nearest local network interface. This value is only used
429 * if the LNET_MD_BULK_HANDLE option is set.
431 struct lnet_handle_md bulk_handle;
434 /* Max Transfer Unit (minimum supported everywhere).
435 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
436 * these limits are system wide and not interface-local. */
437 #define LNET_MTU_BITS 20
438 #define LNET_MTU (1 << LNET_MTU_BITS)
441 * Options for the MD structure. See struct lnet_md::options.
443 #define LNET_MD_OP_PUT (1 << 0)
444 /** See struct lnet_md::options. */
445 #define LNET_MD_OP_GET (1 << 1)
446 /** See struct lnet_md::options. */
447 #define LNET_MD_MANAGE_REMOTE (1 << 2)
448 /* unused (1 << 3) */
449 /** See struct lnet_md::options. */
450 #define LNET_MD_TRUNCATE (1 << 4)
451 /** See struct lnet_md::options. */
452 #define LNET_MD_ACK_DISABLE (1 << 5)
453 /** See struct lnet_md::options. */
454 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
455 /** See struct lnet_md::options. */
456 #define LNET_MD_MAX_SIZE (1 << 7)
457 /** See struct lnet_md::options. */
458 #define LNET_MD_KIOV (1 << 8)
459 /** See struct lnet_md::options. */
460 #define LNET_MD_BULK_HANDLE (1 << 9)
461 /** See struct lnet_md::options. */
462 #define LNET_MD_TRACK_RESPONSE (1 << 10)
463 /** See struct lnet_md::options. */
464 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
465 /** See struct lnet_md::options. */
466 #define LNET_MD_GNILND (1 << 12)
468 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
469 #define LNET_MD_THRESH_INF (-1)
473 /** \addtogroup lnet_eq
477 * Six types of events can be logged in an event queue.
479 enum lnet_event_kind {
480 /** An incoming GET operation has completed on the MD. */
483 * An incoming PUT operation has completed on the MD. The
484 * underlying layers will not alter the memory (on behalf of this
485 * operation) once this event has been logged.
489 * A REPLY operation has completed. This event is logged after the
490 * data (if any) from the REPLY has been written into the MD.
493 /** An acknowledgment has been received. */
496 * An outgoing send (PUT or GET) operation has completed. This event
497 * is logged after the entire buffer has been sent and it is safe for
498 * the caller to reuse the buffer.
501 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
502 * happen even when the message has not yet been put out on wire.
503 * - It's unsafe to assume that in an outgoing GET operation
504 * the LNET_EVENT_SEND event would happen before the
505 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
506 * LNET_EVENT_ACK events in an outgoing PUT operation.
510 * A MD has been unlinked. Note that LNetMDUnlink() does not
511 * necessarily trigger an LNET_EVENT_UNLINK event.
517 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
520 * Information about an event on a MD.
523 /** The identifier (nid, pid) of the target. */
524 struct lnet_processid target;
525 /** The identifier (nid, pid) of the initiator. */
526 struct lnet_processid initiator;
527 /** The source NID on the initiator. */
528 struct lnet_processid source;
530 * The NID of the immediate sender. If the request has been forwarded
531 * by routers, this is the NID of the last hop; otherwise it's the
532 * same as the source.
534 struct lnet_nid sender;
535 /** Indicates the type of the event. */
536 enum lnet_event_kind type;
537 /** The portal table index specified in the request */
538 unsigned int pt_index;
539 /** A copy of the match bits specified in the request. */
541 /** The length (in bytes) specified in the request. */
542 unsigned int rlength;
544 * The length (in bytes) of the data that was manipulated by the
545 * operation. For truncated operations, the manipulated length will be
546 * the number of bytes specified by the MD (possibly with an offset,
547 * see struct lnet_md). For all other operations, the manipulated length
548 * will be the length of the requested operation, i.e. rlength.
550 unsigned int mlength;
552 * The handle to the MD associated with the event. The handle may be
553 * invalid if the MD has been unlinked.
555 struct lnet_handle_md md_handle;
557 * A snapshot of relevant state of the MD immediately after the event
558 * has been processed.
562 unsigned int md_options;
564 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
569 * The message type, to ensure a handler for LNET_EVENT_SEND can
570 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
574 * Indicates the completion status of the operation. It's 0 for
575 * successful operations, otherwise it's an error code.
579 * Indicates whether the MD has been unlinked. Note that:
580 * - An event with unlinked set is the last event on the MD.
581 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
586 * The displacement (in bytes) into the memory region that the
587 * operation used. The offset can be determined by the operation for
588 * a remote managed MD or by the local MD.
589 * \see struct lnet_md::options
593 * The sequence number for this event. Sequence numbers are unique
596 volatile unsigned long sequence;
599 /** \addtogroup lnet_data
603 * Specify whether an acknowledgment should be sent by target when the PUT
604 * operation completes (i.e., when the data has been written to a MD of the
607 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
608 * by which acknowledgments can be disabled for a MD.
611 /** Request an acknowledgment */
613 /** Request that no acknowledgment should be generated. */
618 * UDSP action types. There are two available actions:
619 * 1. PRIORITY - set priority of matching LNet constructs
620 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
622 enum lnet_udsp_action_type {
623 EN_LNET_UDSP_ACTION_NONE = 0,
624 /** assign a priority to matching constructs */
625 EN_LNET_UDSP_ACTION_PRIORITY = 1,
626 /** assign a preferred list of NIDs to matching constructs */
627 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,