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 if (LNET_NID_IS_ANY(nid))
126 return LNET_MKNET(nid->nid_type, __be16_to_cpu(nid->nid_num));
129 static inline void lnet_nid4_to_nid(lnet_nid_t nid4, struct lnet_nid *nid)
131 if (nid4 == LNET_NID_ANY) {
132 /* equal to setting to LNET_ANY_NID */
133 memset(nid, 0xff, sizeof(*nid));
138 nid->nid_type = LNET_NETTYP(LNET_NIDNET(nid4));
139 nid->nid_num = __cpu_to_be16(LNET_NETNUM(LNET_NIDNET(nid4)));
140 nid->nid_addr[0] = __cpu_to_be32(LNET_NIDADDR(nid4));
141 nid->nid_addr[1] = nid->nid_addr[2] = nid->nid_addr[3] = 0;
144 static inline lnet_nid_t lnet_nid_to_nid4(const struct lnet_nid *nid)
146 if (LNET_NID_IS_ANY(nid))
149 return LNET_MKNID(LNET_NID_NET(nid), __be32_to_cpu(nid->nid_addr[0]));
152 static inline int nid_same(const struct lnet_nid *n1,
153 const struct lnet_nid *n2)
155 return n1->nid_size == n2->nid_size &&
156 n1->nid_type == n2->nid_type &&
157 n1->nid_num == n2->nid_num &&
158 n1->nid_addr[0] == n2->nid_addr[0] &&
159 n1->nid_addr[1] == n2->nid_addr[1] &&
160 n1->nid_addr[2] == n2->nid_addr[2] &&
161 n1->nid_addr[3] == n2->nid_addr[3];
164 /* This can be used when we need to hash a nid */
165 static inline unsigned long nidhash(const struct lnet_nid *nid)
168 unsigned long hash = 0;
170 hash ^= LNET_NID_NET(nid);
171 for (i = 0; i < 4; i++)
172 hash ^= nid->nid_addr[i];
176 struct lnet_counters_health {
178 __u32 lch_resend_count;
179 __u32 lch_response_timeout_count;
180 __u32 lch_local_interrupt_count;
181 __u32 lch_local_dropped_count;
182 __u32 lch_local_aborted_count;
183 __u32 lch_local_no_route_count;
184 __u32 lch_local_timeout_count;
185 __u32 lch_local_error_count;
186 __u32 lch_remote_dropped_count;
187 __u32 lch_remote_error_count;
188 __u32 lch_remote_timeout_count;
189 __u32 lch_network_timeout_count;
192 struct lnet_counters {
193 struct lnet_counters_common lct_common;
194 struct lnet_counters_health lct_health;
198 * This is a hard-coded limit on the number of interfaces supported by
199 * the interface bonding implemented by the ksocknal LND. It must be
200 * defined here because it is used in LNet data structures that are
201 * common to all LNDs.
203 #define LNET_INTERFACES_NUM 16
205 /* The minimum number of interfaces per node supported by LNet. */
206 #define LNET_INTERFACES_MIN 16
207 /* The default - arbitrary - value of the lnet_max_interfaces tunable. */
208 #define LNET_INTERFACES_MAX_DEFAULT 200
211 * Objects maintained by the LNet are accessed through handles. Handle types
212 * have names of the form lnet_handle_xx, where xx is one of the two letter
213 * object type codes ('md' for memory descriptor, and
214 * 'me' for match entry). Each type of object is given a unique handle type
215 * to enhance type checking.
217 #define LNET_WIRE_HANDLE_COOKIE_NONE (~0ULL)
219 struct lnet_handle_md {
224 * Invalidate md handle \a h.
226 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
228 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
232 * Check whether handler \a h is invalid.
234 * \return 1 if handle is invalid, 0 if valid.
236 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
238 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
244 struct lnet_process_id {
252 * Global process ID - with large addresses
254 struct lnet_processid {
262 lnet_pid4_to_pid(struct lnet_process_id pid4, struct lnet_processid *pid)
265 lnet_nid4_to_nid(pid4.nid, &pid->nid);
268 static inline struct lnet_process_id
269 lnet_pid_to_pid4(struct lnet_processid *pid)
271 struct lnet_process_id ret;
274 ret.nid = lnet_nid_to_nid4(&pid->nid);
280 /** \addtogroup lnet_me
284 * Specifies whether the match entry or memory descriptor should be unlinked
285 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
293 * Values of the type enum lnet_ins_pos are used to control where a new match
294 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
295 * entry before the current entry or before the head of the list. The value
296 * LNET_INS_AFTER is used to insert the new entry after the current entry
297 * or after the last item in the list.
300 /** insert ME before current position or head of the list */
302 /** insert ME after current position or tail of the list */
304 /** attach ME at tail of local CPU partition ME list */
310 /** \addtogroup lnet_md
313 struct lnet_hdr_nid16 {
314 char _bytes[sizeof(struct lnet_hdr)];
315 } __attribute__((packed));
318 * Event queue handler function type.
320 * The EQ handler runs for each event that is deposited into the EQ. The
321 * handler is supplied with a pointer to the event that triggered the
322 * handler invocation.
324 * The handler must not block, must be reentrant, and must not call any LNet
325 * API functions. It should return as quickly as possible.
328 typedef void (*lnet_handler_t)(struct lnet_event *event);
331 * Defines the visible parts of a memory descriptor. Values of this type
332 * are used to initialize memory descriptors.
336 * Specify the memory region associated with the memory descriptor.
337 * If the options field has:
338 * - LNET_MD_KIOV bit set: The start field points to the starting
339 * address of an array of struct bio_vec and the length field specifies
340 * the number of entries in the array. The length can't be bigger
341 * than LNET_MAX_IOV. The struct bio_vec is used to describe page-based
342 * fragments that are not necessarily mapped in virtal memory.
343 * - Otherwise: The memory region is contiguous. The start field
344 * specifies the starting address for the memory region and the
345 * length field specifies its length.
347 * When the memory region is fragmented, all fragments but the first
348 * one must start on page boundary, and all but the last must end on
354 * Specifies the maximum number of operations that can be performed
355 * on the memory descriptor. An operation is any action that could
356 * possibly generate an event. In the usual case, the threshold value
357 * is decremented for each operation on the MD. When the threshold
358 * drops to zero, the MD becomes inactive and does not respond to
359 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
360 * there is no bound on the number of operations that may be applied
365 * Specifies the largest incoming request that the memory descriptor
366 * should respond to. When the unused portion of a MD (length -
367 * local offset) falls below this value, the MD becomes inactive and
368 * does not respond to further operations. This value is only used
369 * if the LNET_MD_MAX_SIZE option is set.
373 * Specifies the behavior of the memory descriptor. A bitwise OR
374 * of the following values can be used:
375 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
376 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
377 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
378 * region is provided by the incoming request. By default, the
379 * offset is maintained locally. When maintained locally, the
380 * offset is incremented by the length of the request so that
381 * the next operation (PUT or GET) will access the next part of
382 * the memory region. Note that only one offset variable exists
383 * per memory descriptor. If both PUT and GET operations are
384 * performed on a memory descriptor, the offset is updated each time.
385 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
386 * be reduced to match the memory available in the region (determined
387 * by subtracting the offset from the length of the memory region).
388 * By default, if the length in the incoming operation is greater
389 * than the amount of memory available, the operation is rejected.
390 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
391 * incoming PUT operations, even if requested. By default,
392 * acknowledgments are sent for PUT operations that request an
393 * acknowledgment. Acknowledgments are never sent for GET operations.
394 * The data sent in the REPLY serves as an implicit acknowledgment.
395 * - LNET_MD_KIOV: The start and length fields specify an array of
397 * - LNET_MD_MAX_SIZE: The max_size field is valid.
398 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
399 * - LNET_MD_TRACK_RESPONSE: Enable response tracking on this MD
400 * regardless of the value of the lnet_response_tracking param.
401 * - LNET_MD_NO_TRACK_RESPONSE: Disable response tracking on this MD
402 * regardless of the value of the lnet_response_tracking param.
403 * - LNET_MD_GNILND: Disable warning about exceeding LNET_MAX_IOV.
406 * - LNET_MD_KIOV allows for a scatter/gather capability for memory
408 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
409 * region (i.e. sum of all fragment lengths) must not be less than
412 unsigned int options;
414 * A user-specified value that is associated with the memory
415 * descriptor. The value does not need to be a pointer, but must fit
416 * in the space used by a pointer. This value is recorded in events
417 * associated with operations on this MD.
421 * The event handler used to log the operations performed on
422 * the memory region. If this argument is NULL operations
423 * performed on this memory descriptor are not logged.
425 lnet_handler_t handler;
427 * The bulk MD handle which was registered to describe the buffers
428 * either to be used to transfer data to the peer or receive data
429 * from the peer. This allows LNet to properly determine the NUMA
430 * node on which the memory was allocated and use that to select the
431 * nearest local network interface. This value is only used
432 * if the LNET_MD_BULK_HANDLE option is set.
434 struct lnet_handle_md bulk_handle;
437 /* Max Transfer Unit (minimum supported everywhere).
438 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
439 * these limits are system wide and not interface-local. */
440 #define LNET_MTU_BITS 20
441 #define LNET_MTU (1 << LNET_MTU_BITS)
444 * Options for the MD structure. See struct lnet_md::options.
446 #define LNET_MD_OP_PUT (1 << 0)
447 /** See struct lnet_md::options. */
448 #define LNET_MD_OP_GET (1 << 1)
449 /** See struct lnet_md::options. */
450 #define LNET_MD_MANAGE_REMOTE (1 << 2)
451 /* unused (1 << 3) */
452 /** See struct lnet_md::options. */
453 #define LNET_MD_TRUNCATE (1 << 4)
454 /** See struct lnet_md::options. */
455 #define LNET_MD_ACK_DISABLE (1 << 5)
456 /** See struct lnet_md::options. */
457 /* deprecated #define LNET_MD_IOVEC (1 << 6) */
458 /** See struct lnet_md::options. */
459 #define LNET_MD_MAX_SIZE (1 << 7)
460 /** See struct lnet_md::options. */
461 #define LNET_MD_KIOV (1 << 8)
462 /** See struct lnet_md::options. */
463 #define LNET_MD_BULK_HANDLE (1 << 9)
464 /** See struct lnet_md::options. */
465 #define LNET_MD_TRACK_RESPONSE (1 << 10)
466 /** See struct lnet_md::options. */
467 #define LNET_MD_NO_TRACK_RESPONSE (1 << 11)
468 /** See struct lnet_md::options. */
469 #define LNET_MD_GNILND (1 << 12)
470 /** Special page mapping handling */
471 #define LNET_MD_GPU_ADDR (1 << 13)
473 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
474 #define LNET_MD_THRESH_INF (-1)
478 /** \addtogroup lnet_eq
482 * Six types of events can be logged in an event queue.
484 enum lnet_event_kind {
485 /** An incoming GET operation has completed on the MD. */
488 * An incoming PUT operation has completed on the MD. The
489 * underlying layers will not alter the memory (on behalf of this
490 * operation) once this event has been logged.
494 * A REPLY operation has completed. This event is logged after the
495 * data (if any) from the REPLY has been written into the MD.
498 /** An acknowledgment has been received. */
501 * An outgoing send (PUT or GET) operation has completed. This event
502 * is logged after the entire buffer has been sent and it is safe for
503 * the caller to reuse the buffer.
506 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
507 * happen even when the message has not yet been put out on wire.
508 * - It's unsafe to assume that in an outgoing GET operation
509 * the LNET_EVENT_SEND event would happen before the
510 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
511 * LNET_EVENT_ACK events in an outgoing PUT operation.
515 * A MD has been unlinked. Note that LNetMDUnlink() does not
516 * necessarily trigger an LNET_EVENT_UNLINK event.
522 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
525 * Information about an event on a MD.
528 /** The identifier (nid, pid) of the target. */
529 struct lnet_processid target;
530 /** The identifier (nid, pid) of the initiator. */
531 struct lnet_processid initiator;
532 /** The source NID on the initiator. */
533 struct lnet_processid source;
535 * The NID of the immediate sender. If the request has been forwarded
536 * by routers, this is the NID of the last hop; otherwise it's the
537 * same as the source.
539 struct lnet_nid sender;
540 /** Indicates the type of the event. */
541 enum lnet_event_kind type;
542 /** The portal table index specified in the request */
543 unsigned int pt_index;
544 /** A copy of the match bits specified in the request. */
546 /** The length (in bytes) specified in the request. */
547 unsigned int rlength;
549 * The length (in bytes) of the data that was manipulated by the
550 * operation. For truncated operations, the manipulated length will be
551 * the number of bytes specified by the MD (possibly with an offset,
552 * see struct lnet_md). For all other operations, the manipulated length
553 * will be the length of the requested operation, i.e. rlength.
555 unsigned int mlength;
557 * The handle to the MD associated with the event. The handle may be
558 * invalid if the MD has been unlinked.
560 struct lnet_handle_md md_handle;
562 * A snapshot of relevant state of the MD immediately after the event
563 * has been processed.
567 unsigned int md_options;
569 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
574 * The message type, to ensure a handler for LNET_EVENT_SEND can
575 * distinguish between LNET_MSG_GET and LNET_MSG_PUT.
579 * Indicates the completion status of the operation. It's 0 for
580 * successful operations, otherwise it's an error code.
584 * Indicates whether the MD has been unlinked. Note that:
585 * - An event with unlinked set is the last event on the MD.
586 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
591 * The displacement (in bytes) into the memory region that the
592 * operation used. The offset can be determined by the operation for
593 * a remote managed MD or by the local MD.
594 * \see struct lnet_md::options
598 * The sequence number for this event. Sequence numbers are unique
601 volatile unsigned long sequence;
604 /** \addtogroup lnet_data
608 * Specify whether an acknowledgment should be sent by target when the PUT
609 * operation completes (i.e., when the data has been written to a MD of the
612 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
613 * by which acknowledgments can be disabled for a MD.
616 /** Request an acknowledgment */
618 /** Request that no acknowledgment should be generated. */
623 * UDSP action types. There are two available actions:
624 * 1. PRIORITY - set priority of matching LNet constructs
625 * 2. PREFERRED LIST - set preferred list of matching LNet constructs
627 enum lnet_udsp_action_type {
628 EN_LNET_UDSP_ACTION_NONE = 0,
629 /** assign a priority to matching constructs */
630 EN_LNET_UDSP_ACTION_PRIORITY = 1,
631 /** assign a preferred list of NIDs to matching constructs */
632 EN_LNET_UDSP_ACTION_PREFERRED_LIST = 2,