1fi_endpoint(3) Libfabric v1.12.0rc1 fi_endpoint(3)
2
6 fi_endpoint - Fabric endpoint operations
7 fi_endpoint / fi_scalable_ep / fi_passive_ep / fi_close
9 Allocate or close an endpoint.
10 fi_ep_bind
12 Associate an endpoint with hardware resources, such as event
13 queues, completion queues, counters, address vectors, or shared
14 transmit/receive contexts.
15 fi_scalable_ep_bind
17 Associate a scalable endpoint with an address vector
18 fi_pep_bind
20 Associate a passive endpoint with an event queue
21 fi_enable
23 Transitions an active endpoint into an enabled state.
24 fi_cancel
26 Cancel a pending asynchronous data transfer
27 fi_ep_alias
29 Create an alias to the endpoint
30 fi_control
32 Control endpoint operation.
33 fi_getopt / fi_setopt
35 Get or set endpoint options.
36 fi_rx_context / fi_tx_context / fi_srx_context / fi_stx_context
38 Open a transmit or receive context.
39 fi_tc_dscp_set / fi_tc_dscp_get
41 Convert between a DSCP value and a network traffic class
42 fi_rx_size_left / fi_tx_size_left (DEPRECATED)
44 Query the lower bound on how many RX/TX operations may be posted
45 without an operation returning -FI_EAGAIN. This functions have
46 been deprecated and will be removed in a future version of the
47 library.
48
50 #include <rdma/fabric.h>
51 #include <rdma/fi_endpoint.h>
53 int fi_endpoint(struct fid_domain *domain, struct fi_info *info,
55 struct fid_ep **ep, void *context);
56 int fi_scalable_ep(struct fid_domain *domain, struct fi_info *info,
58 struct fid_ep **sep, void *context);
59 int fi_passive_ep(struct fi_fabric *fabric, struct fi_info *info,
61 struct fid_pep **pep, void *context);
62 int fi_tx_context(struct fid_ep *sep, int index,
64 struct fi_tx_attr *attr, struct fid_ep **tx_ep,
65 void *context);
66 int fi_rx_context(struct fid_ep *sep, int index,
68 struct fi_rx_attr *attr, struct fid_ep **rx_ep,
69 void *context);
70 int fi_stx_context(struct fid_domain *domain,
72 struct fi_tx_attr *attr, struct fid_stx **stx,
73 void *context);
74 int fi_srx_context(struct fid_domain *domain,
76 struct fi_rx_attr *attr, struct fid_ep **rx_ep,
77 void *context);
78 int fi_close(struct fid *ep);
80 int fi_ep_bind(struct fid_ep *ep, struct fid *fid, uint64_t flags);
82 int fi_scalable_ep_bind(struct fid_ep *sep, struct fid *fid, uint64_t flags);
84 int fi_pep_bind(struct fid_pep *pep, struct fid *fid, uint64_t flags);
86 int fi_enable(struct fid_ep *ep);
88 int fi_cancel(struct fid_ep *ep, void *context);
90 int fi_ep_alias(struct fid_ep *ep, struct fid_ep **alias_ep, uint64_t flags);
92 int fi_control(struct fid *ep, int command, void *arg);
94 int fi_getopt(struct fid *ep, int level, int optname,
96 void *optval, size_t *optlen);
97 int fi_setopt(struct fid *ep, int level, int optname,
99 const void *optval, size_t optlen);
100 uint32_t fi_tc_dscp_set(uint8_t dscp);
102 uint8_t fi_tc_dscp_get(uint32_t tclass);
104 DEPRECATED ssize_t fi_rx_size_left(struct fid_ep *ep);
106 DEPRECATED ssize_t fi_tx_size_left(struct fid_ep *ep);
108
110 fid On creation, specifies a fabric or access domain. On bind,
111 identifies the event queue, completion queue, counter, or ad‐
112 dress vector to bind to the endpoint. In other cases, it's a
113 fabric identifier of an associated resource.
114 info Details about the fabric interface endpoint to be opened, ob‐
116 tained from fi_getinfo.
117 ep A fabric endpoint.
119 sep A scalable fabric endpoint.
121 pep A passive fabric endpoint.
123 context
125 Context associated with the endpoint or asynchronous operation.
126 index Index to retrieve a specific transmit/receive context.
128 attr Transmit or receive context attributes.
130 flags Additional flags to apply to the operation.
132 command
134 Command of control operation to perform on endpoint.
135 arg Optional control argument.
137 level Protocol level at which the desired option resides.
139 optname
141 The protocol option to read or set.
142 optval The option value that was read or to set.
144 optlen The size of the optval buffer.
146
148 Endpoints are transport level communication portals. There are two
149 types of endpoints: active and passive. Passive endpoints belong to a
150 fabric domain and are most often used to listen for incoming connection
151 requests. However, a passive endpoint may be used to reserve a fabric
152 address that can be granted to an active endpoint. Active endpoints
153 belong to access domains and can perform data transfers.
154 Active endpoints may be connection-oriented or connectionless, and may
156 provide data reliability. The data transfer interfaces -- messages
157 (fi_msg), tagged messages (fi_tagged), RMA (fi_rma), and atomics
158 (fi_atomic) -- are associated with active endpoints. In basic configu‐
159 rations, an active endpoint has transmit and receive queues. In gener‐
160 al, operations that generate traffic on the fabric are posted to the
161 transmit queue. This includes all RMA and atomic operations, along
162 with sent messages and sent tagged messages. Operations that post buf‐
163 fers for receiving incoming data are submitted to the receive queue.
164 Active endpoints are created in the disabled state. They must transi‐
166 tion into an enabled state before accepting data transfer operations,
167 including posting of receive buffers. The fi_enable call is used to
168 transition an active endpoint into an enabled state. The fi_connect
169 and fi_accept calls will also transition an endpoint into the enabled
170 state, if it is not already active.
171 In order to transition an endpoint into an enabled state, it must be
173 bound to one or more fabric resources. An endpoint that will generate
174 asynchronous completions, either through data transfer operations or
175 communication establishment events, must be bound to the appropriate
176 completion queues or event queues, respectively, before being enabled.
177 Additionally, endpoints that use manual progress must be associated
178 with relevant completion queues or event queues in order to drive
179 progress. For endpoints that are only used as the target of RMA or
180 atomic operations, this means binding the endpoint to a completion
181 queue associated with receive processing. Connectionless endpoints
182 must be bound to an address vector.
183 Once an endpoint has been activated, it may be associated with an ad‐
185 dress vector. Receive buffers may be posted to it and calls may be
186 made to connection establishment routines. Connectionless endpoints
187 may also perform data transfers.
188 The behavior of an endpoint may be adjusted by setting its control data
190 and protocol options. This allows the underlying provider to redirect
191 function calls to implementations optimized to meet the desired appli‐
192 cation behavior.
193 If an endpoint experiences a critical error, it will transition back
195 into a disabled state. Critical errors are reported through the event
196 queue associated with the EP. In certain cases, a disabled endpoint
197 may be re-enabled. The ability to transition back into an enabled
198 state is provider specific and depends on the type of error that the
199 endpoint experienced. When an endpoint is disabled as a result of a
200 critical error, all pending operations are discarded.
201 fi_endpoint / fi_passive_ep / fi_scalable_ep
203 fi_endpoint allocates a new active endpoint. fi_passive_ep allocates a
204 new passive endpoint. fi_scalable_ep allocates a scalable endpoint.
205 The properties and behavior of the endpoint are defined based on the
206 provided struct fi_info. See fi_getinfo for additional details on
207 fi_info. fi_info flags that control the operation of an endpoint are
208 defined below. See section SCALABLE ENDPOINTS.
209 If an active endpoint is allocated in order to accept a connection re‐
211 quest, the fi_info parameter must be the same as the fi_info structure
212 provided with the connection request (FI_CONNREQ) event.
213 An active endpoint may acquire the properties of a passive endpoint by
215 setting the fi_info handle field to the passive endpoint fabric de‐
216 scriptor. This is useful for applications that need to reserve the
217 fabric address of an endpoint prior to knowing if the endpoint will be
218 used on the active or passive side of a connection. For example, this
219 feature is useful for simulating socket semantics. Once an active end‐
220 point acquires the properties of a passive endpoint, the passive end‐
221 point is no longer bound to any fabric resources and must no longer be
222 used. The user is expected to close the passive endpoint after opening
223 the active endpoint in order to free up any lingering resources that
224 had been used.
225 fi_close
227 Closes an endpoint and release all resources associated with it.
228 When closing a scalable endpoint, there must be no opened transmit con‐
230 texts, or receive contexts associated with the scalable endpoint. If
231 resources are still associated with the scalable endpoint when attempt‐
232 ing to close, the call will return -FI_EBUSY.
233 Outstanding operations posted to the endpoint when fi_close is called
235 will be discarded. Discarded operations will silently be dropped, with
236 no completions reported. Additionally, a provider may discard previ‐
237 ously completed operations from the associated completion queue(s).
238 The behavior to discard completed operations is provider specific.
239 fi_ep_bind
241 fi_ep_bind is used to associate an endpoint with other allocated re‐
242 sources, such as completion queues, counters, address vectors, event
243 queues, shared contexts, and memory regions. The type of objects that
244 must be bound with an endpoint depend on the endpoint type and its con‐
245 figuration.
246 Passive endpoints must be bound with an EQ that supports connection
248 management events. Connectionless endpoints must be bound to a single
249 address vector. If an endpoint is using a shared transmit and/or re‐
250 ceive context, the shared contexts must be bound to the endpoint. CQs,
251 counters, AV, and shared contexts must be bound to endpoints before
252 they are enabled either explicitly or implicitly.
253 An endpoint must be bound with CQs capable of reporting completions for
255 any asynchronous operation initiated on the endpoint. For example, if
256 the endpoint supports any outbound transfers (sends, RMA, atomics,
257 etc.), then it must be bound to a completion queue that can report
258 transmit completions. This is true even if the endpoint is configured
259 to suppress successful completions, in order that operations that com‐
260 plete in error may be reported to the user.
261 An active endpoint may direct asynchronous completions to different
263 CQs, based on the type of operation. This is specified using
264 fi_ep_bind flags. The following flags may be OR'ed together when bind‐
265 ing an endpoint to a completion domain CQ.
266 FI_RECV
268 Directs the notification of inbound data transfers to the speci‐
269 fied completion queue. This includes received messages. This
270 binding automatically includes FI_REMOTE_WRITE, if applicable to
271 the endpoint.
272 FI_SELECTIVE_COMPLETION
274 By default, data transfer operations write CQ completion entries
275 into the associated completion queue after they have successful‐
276 ly completed. Applications can use this bind flag to selective‐
277 ly enable when completions are generated. If FI_SELECTIVE_COM‐
278 PLETION is specified, data transfer operations will not generate
279 CQ entries for successful completions unless FI_COMPLETION is
280 set as an operational flag for the given operation. Operations
281 that fail asynchronously will still generate completions, even
282 if a completion is not requested. FI_SELECTIVE_COMPLETION must
283 be OR'ed with FI_TRANSMIT and/or FI_RECV flags.
284 When FI_SELECTIVE_COMPLETION is set, the user must determine when a re‐
286 quest that does NOT have FI_COMPLETION set has completed indirectly,
287 usually based on the completion of a subsequent operation or by using
288 completion counters. Use of this flag may improve performance by al‐
289 lowing the provider to avoid writing a CQ completion entry for every
290 operation.
291 See Notes section below for additional information on how this flag in‐
293 teracts with the FI_CONTEXT and FI_CONTEXT2 mode bits.
294 FI_TRANSMIT
296 Directs the completion of outbound data transfer requests to the
297 specified completion queue. This includes send message, RMA,
298 and atomic operations.
299 An endpoint may optionally be bound to a completion counter. Associat‐
301 ing an endpoint with a counter is in addition to binding the EP with a
302 CQ. When binding an endpoint to a counter, the following flags may be
303 specified.
304 FI_READ
306 Increments the specified counter whenever an RMA read, atomic
307 fetch, or atomic compare operation initiated from the endpoint
308 has completed successfully or in error.
309 FI_RECV
311 Increments the specified counter whenever a message is received
312 over the endpoint. Received messages include both tagged and
313 normal message operations.
314 FI_REMOTE_READ
316 Increments the specified counter whenever an RMA read, atomic
317 fetch, or atomic compare operation is initiated from a remote
318 endpoint that targets the given endpoint. Use of this flag re‐
319 quires that the endpoint be created using FI_RMA_EVENT.
320 FI_REMOTE_WRITE
322 Increments the specified counter whenever an RMA write or base
323 atomic operation is initiated from a remote endpoint that tar‐
324 gets the given endpoint. Use of this flag requires that the
325 endpoint be created using FI_RMA_EVENT.
326 FI_SEND
328 Increments the specified counter whenever a message transfer
329 initiated over the endpoint has completed successfully or in er‐
330 ror. Sent messages include both tagged and normal message oper‐
331 ations.
332 FI_WRITE
334 Increments the specified counter whenever an RMA write or base
335 atomic operation initiated from the endpoint has completed suc‐
336 cessfully or in error.
337 An endpoint may only be bound to a single CQ or counter for a given
339 type of operation. For example, a EP may not bind to two counters both
340 using FI_WRITE. Furthermore, providers may limit CQ and counter bind‐
341 ings to endpoints of the same endpoint type (DGRAM, MSG, RDM, etc.).
342 fi_scalable_ep_bind
344 fi_scalable_ep_bind is used to associate a scalable endpoint with an
345 address vector. See section on SCALABLE ENDPOINTS. A scalable end‐
346 point has a single transport level address and can support multiple
347 transmit and receive contexts. The transmit and receive contexts share
348 the transport-level address. Address vectors that are bound to scal‐
349 able endpoints are implicitly bound to any transmit or receive contexts
350 created using the scalable endpoint.
351 fi_enable
353 This call transitions the endpoint into an enabled state. An endpoint
354 must be enabled before it may be used to perform data transfers. En‐
355 abling an endpoint typically results in hardware resources being as‐
356 signed to it. Endpoints making use of completion queues, counters,
357 event queues, and/or address vectors must be bound to them before being
358 enabled.
359 Calling connect or accept on an endpoint will implicitly enable an end‐
361 point if it has not already been enabled.
362 fi_enable may also be used to re-enable an endpoint that has been dis‐
364 abled as a result of experiencing a critical error. Applications
365 should check the return value from fi_enable to see if a disabled end‐
366 point has successfully be re-enabled.
367 fi_cancel
369 fi_cancel attempts to cancel an outstanding asynchronous operation.
370 Canceling an operation causes the fabric provider to search for the op‐
371 eration and, if it is still pending, complete it as having been can‐
372 celed. An error queue entry will be available in the associated error
373 queue with error code FI_ECANCELED. On the other hand, if the opera‐
374 tion completed before the call to fi_cancel, then the completion status
375 of that operation will be available in the associated completion queue.
376 No specific entry related to fi_cancel itself will be posted.
377 Cancel uses the context parameter associated with an operation to iden‐
379 tify the request to cancel. Operations posted without a valid context
380 parameter -- either no context parameter is specified or the context
381 value was ignored by the provider -- cannot be canceled. If multiple
382 outstanding operations match the context parameter, only one will be
383 canceled. In this case, the operation which is canceled is provider
384 specific. The cancel operation is asynchronous, but will complete
385 within a bounded period of time.
386 fi_ep_alias
388 This call creates an alias to the specified endpoint. Conceptually, an
389 endpoint alias provides an alternate software path from the application
390 to the underlying provider hardware. An alias EP differs from its par‐
391 ent endpoint only by its default data transfer flags. For example, an
392 alias EP may be configured to use a different completion mode. By de‐
393 fault, an alias EP inherits the same data transfer flags as the parent
394 endpoint. An application can use fi_control to modify the alias EP op‐
395 erational flags.
396 When allocating an alias, an application may configure either the
398 transmit or receive operational flags. This avoids needing a separate
399 call to fi_control to set those flags. The flags passed to fi_ep_alias
400 must include FI_TRANSMIT or FI_RECV (not both) with other operational
401 flags OR'ed in. This will override the transmit or receive flags, re‐
402 spectively, for operations posted through the alias endpoint. All al‐
403 located aliases must be closed for the underlying endpoint to be re‐
404 leased.
405 fi_control
407 The control operation is used to adjust the default behavior of an end‐
408 point. It allows the underlying provider to redirect function calls to
409 implementations optimized to meet the desired application behavior. As
410 a result, calls to fi_ep_control must be serialized against all other
411 calls to an endpoint.
412 The base operation of an endpoint is selected during creation using
414 struct fi_info. The following control commands and arguments may be
415 assigned to an endpoint.
416 **FI_BACKLOG - int *value**
418 This option only applies to passive endpoints. It is used to
419 set the connection request backlog for listening endpoints.
420 **FI_GETOPSFLAG -- uint64_t *flags**
422 Used to retrieve the current value of flags associated with the
423 data transfer operations initiated on the endpoint. The control
424 argument must include FI_TRANSMIT or FI_RECV (not both) flags to
425 indicate the type of data transfer flags to be returned. See
426 below for a list of control flags.
427 FI_GETWAIT -- void **
429 This command allows the user to retrieve the file descriptor as‐
430 sociated with a socket endpoint. The fi_control arg parameter
431 should be an address where a pointer to the returned file de‐
432 scriptor will be written. See fi_eq.3 for addition details us‐
433 ing fi_control with FI_GETWAIT. The file descriptor may be used
434 for notification that the endpoint is ready to send or receive
435 data.
436 **FI_SETOPSFLAG -- uint64_t *flags**
438 Used to change the data transfer operation flags associated with
439 an endpoint. The control argument must include FI_TRANSMIT or
440 FI_RECV (not both) to indicate the type of data transfer that
441 the flags should apply to, with other flags OR'ed in. The given
442 flags will override the previous transmit and receive attributes
443 that were set when the endpoint was created. Valid control
444 flags are defined below.
445 fi_getopt / fi_setopt
447 Endpoint protocol operations may be retrieved using fi_getopt or set
448 using fi_setopt. Applications specify the level that a desired option
449 exists, identify the option, and provide input/output buffers to get or
450 set the option. fi_setopt provides an application a way to adjust
451 low-level protocol and implementation specific details of an endpoint.
452 The following option levels and option names and parameters are de‐
454 fined.
455 FI_OPT_ENDPOINT · .RS 2
457 FI_OPT_BUFFERED_LIMIT - size_t
459 Defines the maximum size of a buffered message that will be re‐
460 ported to users as part of a receive completion when the
461 FI_BUFFERED_RECV mode is enabled on an endpoint.
462 fi_getopt() will return the currently configured threshold, or the
464 provider's default threshold if one has not be set by the application.
465 fi_setopt() allows an application to configure the threshold. If the
466 provider cannot support the requested threshold, it will fail the
467 fi_setopt() call with FI_EMSGSIZE. Calling fi_setopt() with the
468 threshold set to SIZE_MAX will set the threshold to the maximum sup‐
469 ported by the provider. fi_getopt() can then be used to retrieve the
470 set size.
471 In most cases, the sending and receiving endpoints must be configured
473 to use the same threshold value, and the threshold must be set prior to
474 enabling the endpoint. · .RS 2
475 FI_OPT_BUFFERED_MIN - size_t
477 Defines the minimum size of a buffered message that will be re‐
478 ported. Applications would set this to a size that's big enough
479 to decide whether to discard or claim a buffered receive or when
480 to claim a buffered receive on getting a buffered receive com‐
481 pletion. The value is typically used by a provider when sending
482 a rendezvous protocol request where it would send at least
483 FI_OPT_BUFFERED_MIN bytes of application data along with it. A
484 smaller sized rendezvous protocol message usually results in
485 better latency for the overall transfer of a large message.
486 · .RS 2
487 FI_OPT_CM_DATA_SIZE - size_t
489 Defines the size of available space in CM messages for user-de‐
490 fined data. This value limits the amount of data that applica‐
491 tions can exchange between peer endpoints using the fi_connect,
492 fi_accept, and fi_reject operations. The size returned is de‐
493 pendent upon the properties of the endpoint, except in the case
494 of passive endpoints, in which the size reflects the maximum
495 size of the data that may be present as part of a connection re‐
496 quest event. This option is read only.
497 · .RS 2
498 FI_OPT_MIN_MULTI_RECV - size_t
500 Defines the minimum receive buffer space available when the re‐
501 ceive buffer is released by the provider (see FI_MULTI_RECV).
502 Modifying this value is only guaranteed to set the minimum buf‐
503 fer space needed on receives posted after the value has been
504 changed. It is recommended that applications that want to over‐
505 ride the default MIN_MULTI_RECV value set this option before en‐
506 abling the corresponding endpoint.
507 fi_tc_dscp_set
509 This call converts a DSCP defined value into a libfabric traffic class
510 value. It should be used when assigning a DSCP value when setting the
511 tclass field in either domain or endpoint attributes
512 fi_tc_dscp_get
514 This call returns the DSCP value associated with the tclass field for
515 the domain or endpoint attributes.
516 fi_rx_size_left (DEPRECATED)
518 This function has been deprecated and will be removed in a future ver‐
519 sion of the library. It may not be supported by all providers.
520 The fi_rx_size_left call returns a lower bound on the number of receive
522 operations that may be posted to the given endpoint without that opera‐
523 tion returning -FI_EAGAIN. Depending on the specific details of the
524 subsequently posted receive operations (e.g., number of iov entries,
525 which receive function is called, etc.), it may be possible to post
526 more receive operations than originally indicated by fi_rx_size_left.
527 fi_tx_size_left (DEPRECATED)
529 This function has been deprecated and will be removed in a future ver‐
530 sion of the library. It may not be supported by all providers.
531 The fi_tx_size_left call returns a lower bound on the number of trans‐
533 mit operations that may be posted to the given endpoint without that
534 operation returning -FI_EAGAIN. Depending on the specific details of
535 the subsequently posted transmit operations (e.g., number of iov en‐
536 tries, which transmit function is called, etc.), it may be possible to
537 post more transmit operations than originally indicated by
538 fi_tx_size_left.
539
541 The fi_ep_attr structure defines the set of attributes associated with
542 an endpoint. Endpoint attributes may be further refined using the
543 transmit and receive context attributes as shown below.
544 struct fi_ep_attr {
546 enum fi_ep_type type;
547 uint32_t protocol;
548 uint32_t protocol_version;
549 size_t max_msg_size;
550 size_t msg_prefix_size;
551 size_t max_order_raw_size;
552 size_t max_order_war_size;
553 size_t max_order_waw_size;
554 uint64_t mem_tag_format;
555 size_t tx_ctx_cnt;
556 size_t rx_ctx_cnt;
557 size_t auth_key_size;
558 uint8_t *auth_key;
559 };
560 type - Endpoint Type
562 If specified, indicates the type of fabric interface communication de‐
563 sired. Supported types are:
564 FI_EP_DGRAM
566 Supports a connectionless, unreliable datagram communication.
567 Message boundaries are maintained, but the maximum message size
568 may be limited to the fabric MTU. Flow control is not guaran‐
569 teed.
570 FI_EP_MSG
572 Provides a reliable, connection-oriented data transfer service
573 with flow control that maintains message boundaries.
574 FI_EP_RDM
576 Reliable datagram message. Provides a reliable, connectionless
577 data transfer service with flow control that maintains message
578 boundaries.
579 FI_EP_SOCK_DGRAM
581 A connectionless, unreliable datagram endpoint with UDP sock‐
582 et-like semantics. FI_EP_SOCK_DGRAM is most useful for applica‐
583 tions designed around using UDP sockets. See the SOCKET END‐
584 POINT section for additional details and restrictions that apply
585 to datagram socket endpoints.
586 FI_EP_SOCK_STREAM
588 Data streaming endpoint with TCP socket-like semantics. Pro‐
589 vides a reliable, connection-oriented data transfer service that
590 does not maintain message boundaries. FI_EP_SOCK_STREAM is most
591 useful for applications designed around using TCP sockets. See
592 the SOCKET ENDPOINT section for additional details and restric‐
593 tions that apply to stream endpoints.
594 FI_EP_UNSPEC
596 The type of endpoint is not specified. This is usually provided
597 as input, with other attributes of the endpoint or the provider
598 selecting the type.
599 Protocol
601 Specifies the low-level end to end protocol employed by the provider.
602 A matching protocol must be used by communicating endpoints to ensure
603 interoperability. The following protocol values are defined. Provider
604 specific protocols are also allowed. Provider specific protocols will
605 be indicated by having the upper bit of the protocol value set to one.
606 FI_PROTO_GNI
608 Protocol runs over Cray GNI low-level interface.
609 FI_PROTO_IB_RDM
611 Reliable-datagram protocol implemented over InfiniBand reli‐
612 able-connected queue pairs.
613 FI_PROTO_IB_UD
615 The protocol runs over Infiniband unreliable datagram queue
616 pairs.
617 FI_PROTO_IWARP
619 The protocol runs over the Internet wide area RDMA protocol
620 transport.
621 FI_PROTO_IWARP_RDM
623 Reliable-datagram protocol implemented over iWarp reliable-con‐
624 nected queue pairs.
625 FI_PROTO_NETWORKDIRECT
627 Protocol runs over Microsoft NetworkDirect service provider in‐
628 terface. This adds reliable-datagram semantics over the Net‐
629 workDirect connection- oriented endpoint semantics.
630 FI_PROTO_PSMX
632 The protocol is based on an Intel proprietary protocol known as
633 PSM, performance scaled messaging. PSMX is an extended version
634 of the PSM protocol to support the libfabric interfaces.
635 FI_PROTO_PSMX2
637 The protocol is based on an Intel proprietary protocol known as
638 PSM2, performance scaled messaging version 2. PSMX2 is an ex‐
639 tended version of the PSM2 protocol to support the libfabric in‐
640 terfaces.
641 FI_PROTO_RDMA_CM_IB_RC
643 The protocol runs over Infiniband reliable-connected queue
644 pairs, using the RDMA CM protocol for connection establishment.
645 FI_PROTO_RXD
647 Reliable-datagram protocol implemented over datagram endpoints.
648 RXD is a libfabric utility component that adds RDM endpoint se‐
649 mantics over DGRAM endpoint semantics.
650 FI_PROTO_RXM
652 Reliable-datagram protocol implemented over message endpoints.
653 RXM is a libfabric utility component that adds RDM endpoint se‐
654 mantics over MSG endpoint semantics.
655 FI_PROTO_SOCK_TCP
657 The protocol is layered over TCP packets.
658 FI_PROTO_UDP
660 The protocol sends and receives UDP datagrams. For example, an
661 endpoint using FI_PROTO_UDP will be able to communicate with a
662 remote peer that is using Berkeley SOCK_DGRAM sockets using IP‐
663 PROTO_UDP.
664 FI_PROTO_UNSPEC
666 The protocol is not specified. This is usually provided as in‐
667 put, with other attributes of the socket or the provider select‐
668 ing the actual protocol.
669 protocol_version - Protocol Version
671 Identifies which version of the protocol is employed by the provider.
672 The protocol version allows providers to extend an existing protocol,
673 by adding support for additional features or functionality for example,
674 in a backward compatible manner. Providers that support different ver‐
675 sions of the same protocol should inter-operate, but only when using
676 the capabilities defined for the lesser version.
677 max_msg_size - Max Message Size
679 Defines the maximum size for an application data transfer as a single
680 operation.
681 msg_prefix_size - Message Prefix Size
683 Specifies the size of any required message prefix buffer space. This
684 field will be 0 unless the FI_MSG_PREFIX mode is enabled. If msg_pre‐
685 fix_size is > 0 the specified value will be a multiple of 8-bytes.
686 Max RMA Ordered Size
688 The maximum ordered size specifies the delivery order of transport data
689 into target memory for RMA and atomic operations. Data ordering is
690 separate, but dependent on message ordering (defined below). Data or‐
691 dering is unspecified where message order is not defined.
692 Data ordering refers to the access of target memory by subsequent oper‐
694 ations. When back to back RMA read or write operations access the same
695 registered memory location, data ordering indicates whether the second
696 operation reads or writes the target memory after the first operation
697 has completed. Because RMA ordering applies between two operations,
698 and not within a single data transfer, ordering is defined per byte-ad‐
699 dressable memory location. I.e. ordering specifies whether location X
700 is accessed by the second operation after the first operation. Nothing
701 is implied about the completion of the first operation before the sec‐
702 ond operation is initiated.
703 In order to support large data transfers being broken into multiple
705 packets and sent using multiple paths through the fabric, data ordering
706 may be limited to transfers of a specific size or less. Providers
707 specify when data ordering is maintained through the following values.
708 Note that even if data ordering is not maintained, message ordering may
709 be.
710 max_order_raw_size
712 Read after write size. If set, an RMA or atomic read operation
713 issued after an RMA or atomic write operation, both of which are
714 smaller than the size, will be ordered. Where the target memory
715 locations overlap, the RMA or atomic read operation will see the
716 results of the previous RMA or atomic write.
717 max_order_war_size
719 Write after read size. If set, an RMA or atomic write operation
720 issued after an RMA or atomic read operation, both of which are
721 smaller than the size, will be ordered. The RMA or atomic read
722 operation will see the initial value of the target memory loca‐
723 tion before a subsequent RMA or atomic write updates the value.
724 max_order_waw_size
726 Write after write size. If set, an RMA or atomic write opera‐
727 tion issued after an RMA or atomic write operation, both of
728 which are smaller than the size, will be ordered. The target
729 memory location will reflect the results of the second RMA or
730 atomic write.
731 An order size value of 0 indicates that ordering is not guaranteed. A
733 value of -1 guarantees ordering for any data size.
734 mem_tag_format - Memory Tag Format
736 The memory tag format is a bit array used to convey the number of
737 tagged bits supported by a provider. Additionally, it may be used to
738 divide the bit array into separate fields. The mem_tag_format option‐
739 ally begins with a series of bits set to 0, to signify bits which are
740 ignored by the provider. Following the initial prefix of ignored bits,
741 the array will consist of alternating groups of bits set to all 1's or
742 all 0's. Each group of bits corresponds to a tagged field. The impli‐
743 cation of defining a tagged field is that when a mask is applied to the
744 tagged bit array, all bits belonging to a single field will either be
745 set to 1 or 0, collectively.
746 For example, a mem_tag_format of 0x30FF indicates support for 14 tagged
748 bits, separated into 3 fields. The first field consists of 2-bits, the
749 second field 4-bits, and the final field 8-bits. Valid masks for such
750 a tagged field would be a bitwise OR'ing of zero or more of the follow‐
751 ing values: 0x3000, 0x0F00, and 0x00FF. The provider may not validate
752 the mask provided by the application for performance reasons.
753 By identifying fields within a tag, a provider may be able to optimize
755 their search routines. An application which requests tag fields must
756 provide tag masks that either set all mask bits corresponding to a
757 field to all 0 or all 1. When negotiating tag fields, an application
758 can request a specific number of fields of a given size. A provider
759 must return a tag format that supports the requested number of fields,
760 with each field being at least the size requested, or fail the request.
761 A provider may increase the size of the fields. When reporting comple‐
762 tions (see FI_CQ_FORMAT_TAGGED), it is not guaranteed that the provider
763 would clear out any unsupported tag bits in the tag field of the com‐
764 pletion entry.
765 It is recommended that field sizes be ordered from smallest to largest.
767 A generic, unstructured tag and mask can be achieved by requesting a
768 bit array consisting of alternating 1's and 0's.
769 tx_ctx_cnt - Transmit Context Count
771 Number of transmit contexts to associate with the endpoint. If not
772 specified (0), 1 context will be assigned if the endpoint supports out‐
773 bound transfers. Transmit contexts are independent transmit queues
774 that may be separately configured. Each transmit context may be bound
775 to a separate CQ, and no ordering is defined between contexts. Addi‐
776 tionally, no synchronization is needed when accessing contexts in par‐
777 allel.
778 If the count is set to the value FI_SHARED_CONTEXT, the endpoint will
780 be configured to use a shared transmit context, if supported by the
781 provider. Providers that do not support shared transmit contexts will
782 fail the request.
783 See the scalable endpoint and shared contexts sections for additional
785 details.
786 rx_ctx_cnt - Receive Context Count
788 Number of receive contexts to associate with the endpoint. If not
789 specified, 1 context will be assigned if the endpoint supports inbound
790 transfers. Receive contexts are independent processing queues that may
791 be separately configured. Each receive context may be bound to a sepa‐
792 rate CQ, and no ordering is defined between contexts. Additionally, no
793 synchronization is needed when accessing contexts in parallel.
794 If the count is set to the value FI_SHARED_CONTEXT, the endpoint will
796 be configured to use a shared receive context, if supported by the
797 provider. Providers that do not support shared receive contexts will
798 fail the request.
799 See the scalable endpoint and shared contexts sections for additional
801 details.
802 auth_key_size - Authorization Key Length
804 The length of the authorization key in bytes. This field will be 0 if
805 authorization keys are not available or used. This field is ignored
806 unless the fabric is opened with API version 1.5 or greater.
807 auth_key - Authorization Key
809 If supported by the fabric, an authorization key (a.k.a. job key) to
810 associate with the endpoint. An authorization key is used to limit
811 communication between endpoints. Only peer endpoints that are pro‐
812 grammed to use the same authorization key may communicate. Authoriza‐
813 tion keys are often used to implement job keys, to ensure that process‐
814 es running in different jobs do not accidentally cross traffic. The
815 domain authorization key will be used if auth_key_size is set to 0.
816 This field is ignored unless the fabric is opened with API version 1.5
817 or greater.
818
820 Attributes specific to the transmit capabilities of an endpoint are
821 specified using struct fi_tx_attr.
822 struct fi_tx_attr {
824 uint64_t caps;
825 uint64_t mode;
826 uint64_t op_flags;
827 uint64_t msg_order;
828 uint64_t comp_order;
829 size_t inject_size;
830 size_t size;
831 size_t iov_limit;
832 size_t rma_iov_limit;
833 uint32_t tclass;
834 };
835 caps - Capabilities
837 The requested capabilities of the context. The capabilities must be a
838 subset of those requested of the associated endpoint. See the CAPABIL‐
839 ITIES section of fi_getinfo(3) for capability details. If the caps
840 field is 0 on input to fi_getinfo(3), the applicable capability bits
841 from the fi_info structure will be used.
842 The following capabilities apply to the transmit attributes: FI_MSG,
844 FI_RMA, FI_TAGGED, FI_ATOMIC, FI_READ, FI_WRITE, FI_SEND, FI_HMEM,
845 FI_TRIGGER, FI_FENCE, FI_MULTICAST, FI_RMA_PMEM, FI_NAMED_RX_CTX, and
846 FI_COLLECTIVE.
847 Many applications will be able to ignore this field and rely solely on
849 the fi_info::caps field. Use of this field provides fine grained con‐
850 trol over the transmit capabilities associated with an endpoint. It is
851 useful when handling scalable endpoints, with multiple transmit con‐
852 texts, for example, and allows configuring a specific transmit context
853 with fewer capabilities than that supported by the endpoint or other
854 transmit contexts.
855 mode
857 The operational mode bits of the context. The mode bits will be a sub‐
858 set of those associated with the endpoint. See the MODE section of
859 fi_getinfo(3) for details. A mode value of 0 will be ignored on input
860 to fi_getinfo(3), with the mode value of the fi_info structure used in‐
861 stead. On return from fi_getinfo(3), the mode will be set only to
862 those constraints specific to transmit operations.
863 op_flags - Default transmit operation flags
865 Flags that control the operation of operations submitted against the
866 context. Applicable flags are listed in the Operation Flags section.
867 msg_order - Message Ordering
869 Message ordering refers to the order in which transport layer headers
870 (as viewed by the application) are identified and processed. Relaxed
871 message order enables data transfers to be sent and received out of or‐
872 der, which may improve performance by utilizing multiple paths through
873 the fabric from the initiating endpoint to a target endpoint. Message
874 order applies only between a single source and destination endpoint
875 pair. Ordering between different target endpoints is not defined.
876 Message order is determined using a set of ordering bits. Each set bit
878 indicates that ordering is maintained between data transfers of the
879 specified type. Message order is defined for [read | write | send] op‐
880 erations submitted by an application after [read | write | send] opera‐
881 tions.
882 Message ordering only applies to the end to end transmission of trans‐
884 port headers. Message ordering is necessary, but does not guarantee,
885 the order in which message data is sent or received by the transport
886 layer. Message ordering requires matching ordering semantics on the
887 receiving side of a data transfer operation in order to guarantee that
888 ordering is met.
889 FI_ORDER_ATOMIC_RAR
891 Atomic read after read. If set, atomic fetch operations are
892 transmitted in the order submitted relative to other atomic
893 fetch operations. If not set, atomic fetches may be transmitted
894 out of order from their submission.
895 FI_ORDER_ATOMIC_RAW
897 Atomic read after write. If set, atomic fetch operations are
898 transmitted in the order submitted relative to atomic update op‐
899 erations. If not set, atomic fetches may be transmitted ahead
900 of atomic updates.
901 FI_ORDER_ATOMIC_WAR
903 RMA write after read. If set, atomic update operations are
904 transmitted in the order submitted relative to atomic fetch op‐
905 erations. If not set, atomic updates may be transmitted ahead
906 of atomic fetches.
907 FI_ORDER_ATOMIC_WAW
909 RMA write after write. If set, atomic update operations are
910 transmitted in the order submitted relative to other atomic up‐
911 date operations. If not atomic updates may be transmitted out
912 of order from their submission.
913 FI_ORDER_NONE
915 No ordering is specified. This value may be used as input in
916 order to obtain the default message order supported by the
917 provider. FI_ORDER_NONE is an alias for the value 0.
918 FI_ORDER_RAR
920 Read after read. If set, RMA and atomic read operations are
921 transmitted in the order submitted relative to other RMA and
922 atomic read operations. If not set, RMA and atomic reads may be
923 transmitted out of order from their submission.
924 FI_ORDER_RAS
926 Read after send. If set, RMA and atomic read operations are
927 transmitted in the order submitted relative to message send op‐
928 erations, including tagged sends. If not set, RMA and atomic
929 reads may be transmitted ahead of sends.
930 FI_ORDER_RAW
932 Read after write. If set, RMA and atomic read operations are
933 transmitted in the order submitted relative to RMA and atomic
934 write operations. If not set, RMA and atomic reads may be
935 transmitted ahead of RMA and atomic writes.
936 FI_ORDER_RMA_RAR
938 RMA read after read. If set, RMA read operations are transmit‐
939 ted in the order submitted relative to other RMA read opera‐
940 tions. If not set, RMA reads may be transmitted out of order
941 from their submission.
942 FI_ORDER_RMA_RAW
944 RMA read after write. If set, RMA read operations are transmit‐
945 ted in the order submitted relative to RMA write operations. If
946 not set, RMA reads may be transmitted ahead of RMA writes.
947 FI_ORDER_RMA_WAR
949 RMA write after read. If set, RMA write operations are trans‐
950 mitted in the order submitted relative to RMA read operations.
951 If not set, RMA writes may be transmitted ahead of RMA reads.
952 FI_ORDER_RMA_WAW
954 RMA write after write. If set, RMA write operations are trans‐
955 mitted in the order submitted relative to other RMA write opera‐
956 tions. If not set, RMA writes may be transmitted out of order
957 from their submission.
958 FI_ORDER_SAR
960 Send after read. If set, message send operations, including
961 tagged sends, are transmitted in order submitted relative to RMA
962 and atomic read operations. If not set, message sends may be
963 transmitted ahead of RMA and atomic reads.
964 FI_ORDER_SAS
966 Send after send. If set, message send operations, including
967 tagged sends, are transmitted in the order submitted relative to
968 other message send. If not set, message sends may be transmit‐
969 ted out of order from their submission.
970 FI_ORDER_SAW
972 Send after write. If set, message send operations, including
973 tagged sends, are transmitted in order submitted relative to RMA
974 and atomic write operations. If not set, message sends may be
975 transmitted ahead of RMA and atomic writes.
976 FI_ORDER_WAR
978 Write after read. If set, RMA and atomic write operations are
979 transmitted in the order submitted relative to RMA and atomic
980 read operations. If not set, RMA and atomic writes may be
981 transmitted ahead of RMA and atomic reads.
982 FI_ORDER_WAS
984 Write after send. If set, RMA and atomic write operations are
985 transmitted in the order submitted relative to message send op‐
986 erations, including tagged sends. If not set, RMA and atomic
987 writes may be transmitted ahead of sends.
988 FI_ORDER_WAW
990 Write after write. If set, RMA and atomic write operations are
991 transmitted in the order submitted relative to other RMA and
992 atomic write operations. If not set, RMA and atomic writes may
993 be transmitted out of order from their submission.
994 comp_order - Completion Ordering
996 Completion ordering refers to the order in which completed requests are
997 written into the completion queue. Completion ordering is similar to
998 message order. Relaxed completion order may enable faster reporting of
999 completed transfers, allow acknowledgments to be sent over different
1000 fabric paths, and support more sophisticated retry mechanisms. This
1001 can result in lower-latency completions, particularly when using con‐
1002 nectionless endpoints. Strict completion ordering may require that
1003 providers queue completed operations or limit available optimizations.
1004 For transmit requests, completion ordering depends on the endpoint com‐
1006 munication type. For unreliable communication, completion ordering ap‐
1007 plies to all data transfer requests submitted to an endpoint. For re‐
1008 liable communication, completion ordering only applies to requests that
1009 target a single destination endpoint. Completion ordering of requests
1010 that target different endpoints over a reliable transport is not de‐
1011 fined.
1012 Applications should specify the completion ordering that they support
1014 or require. Providers should return the completion order that they ac‐
1015 tually provide, with the constraint that the returned ordering is
1016 stricter than that specified by the application. Supported completion
1017 order values are:
1018 FI_ORDER_NONE
1020 No ordering is defined for completed operations. Requests sub‐
1021 mitted to the transmit context may complete in any order.
1022 FI_ORDER_STRICT
1024 Requests complete in the order in which they are submitted to
1025 the transmit context.
1026 inject_size
1028 The requested inject operation size (see the FI_INJECT flag) that the
1029 context will support. This is the maximum size data transfer that can
1030 be associated with an inject operation (such as fi_inject) or may be
1031 used with the FI_INJECT data transfer flag.
1032 size
1034 The size of the transmit context. The mapping of the size value to re‐
1035 sources is provider specific, but it is directly related to the number
1036 of command entries allocated for the endpoint. A smaller size value
1037 consumes fewer hardware and software resources, while a larger size al‐
1038 lows queuing more transmit requests.
1039 While the size attribute guides the size of underlying endpoint trans‐
1041 mit queue, there is not necessarily a one-to-one mapping between a
1042 transmit operation and a queue entry. A single transmit operation may
1043 consume multiple queue entries; for example, one per scatter-gather en‐
1044 try. Additionally, the size field is intended to guide the allocation
1045 of the endpoint's transmit context. Specifically, for connectionless
1046 endpoints, there may be lower-level queues use to track communication
1047 on a per peer basis. The sizes of any lower-level queues may only be
1048 significantly smaller than the endpoint's transmit size, in order to
1049 reduce resource utilization.
1050 iov_limit
1052 This is the maximum number of IO vectors (scatter-gather elements) that
1053 a single posted operation may reference.
1054 rma_iov_limit
1056 This is the maximum number of RMA IO vectors (scatter-gather elements)
1057 that an RMA or atomic operation may reference. The rma_iov_limit cor‐
1058 responds to the rma_iov_count values in RMA and atomic operations. See
1059 struct fi_msg_rma and struct fi_msg_atomic in fi_rma.3 and fi_atomic.3,
1060 for additional details. This limit applies to both the number of RMA
1061 IO vectors that may be specified when initiating an operation from the
1062 local endpoint, as well as the maximum number of IO vectors that may be
1063 carried in a single request from a remote endpoint.
1064 Traffic Class (tclass)
1066 Traffic classes can be a differentiated services code point (DSCP) val‐
1067 ue, one of the following defined labels, or a provider-specific defini‐
1068 tion. If tclass is unset or set to FI_TC_UNSPEC, the endpoint will use
1069 the default traffic class associated with the domain.
1070 FI_TC_BEST_EFFORT
1072 This is the default in the absence of any other local or fabric
1073 configuration. This class carries the traffic for a number of
1074 applications executing concurrently over the same network infra‐
1075 structure. Even though it is shared, network capacity and re‐
1076 source allocation are distributed fairly across the applica‐
1077 tions.
1078 FI_TC_BULK_DATA
1080 This class is intended for large data transfers associated with
1081 I/O and is present to separate sustained I/O transfers from oth‐
1082 er application inter-process communications.
1083 FI_TC_DEDICATED_ACCESS
1085 This class operates at the highest priority, except the manage‐
1086 ment class. It carries a high bandwidth allocation, minimum la‐
1087 tency targets, and the highest scheduling and arbitration prior‐
1088 ity.
1089 FI_TC_LOW_LATENCY
1091 This class supports low latency, low jitter data patterns typi‐
1092 cally caused by transactional data exchanges, barrier synchro‐
1093 nizations, and collective operations that are typical of HPC ap‐
1094 plications. This class often requires maximum tolerable laten‐
1095 cies that data transfers must achieve for correct or performance
1096 operations. Fulfillment of such requests in this class will
1097 typically require accompanying bandwidth and message size limi‐
1098 tations so as not to consume excessive bandwidth at high priori‐
1099 ty.
1100 FI_TC_NETWORK_CTRL
1102 This class is intended for traffic directly related to fabric
1103 (network) management, which is critical to the correct operation
1104 of the network. Its use is typically restricted to privileged
1105 network management applications.
1106 FI_TC_SCAVENGER
1108 This class is used for data that is desired but does not have
1109 strict delivery requirements, such as in-band network or appli‐
1110 cation level monitoring data. Use of this class indicates that
1111 the traffic is considered lower priority and should not inter‐
1112 fere with higher priority workflows.
1113 fi_tc_dscp_set / fi_tc_dscp_get
1115 DSCP values are supported via the DSCP get and set functions.
1116 The definitions for DSCP values are outside the scope of libfab‐
1117 ric. See the fi_tc_dscp_set and fi_tc_dscp_get function defini‐
1118 tions for details on their use.
1119
1121 Attributes specific to the receive capabilities of an endpoint are
1122 specified using struct fi_rx_attr.
1123 struct fi_rx_attr {
1125 uint64_t caps;
1126 uint64_t mode;
1127 uint64_t op_flags;
1128 uint64_t msg_order;
1129 uint64_t comp_order;
1130 size_t total_buffered_recv;
1131 size_t size;
1132 size_t iov_limit;
1133 };
1134 caps - Capabilities
1136 The requested capabilities of the context. The capabilities must be a
1137 subset of those requested of the associated endpoint. See the CAPABIL‐
1138 ITIES section if fi_getinfo(3) for capability details. If the caps
1139 field is 0 on input to fi_getinfo(3), the applicable capability bits
1140 from the fi_info structure will be used.
1141 The following capabilities apply to the receive attributes: FI_MSG,
1143 FI_RMA, FI_TAGGED, FI_ATOMIC, FI_REMOTE_READ, FI_REMOTE_WRITE, FI_RECV,
1144 FI_HMEM, FI_TRIGGER, FI_RMA_PMEM, FI_DIRECTED_RECV, FI_VARIABLE_MSG,
1145 FI_MULTI_RECV, FI_SOURCE, FI_RMA_EVENT, FI_SOURCE_ERR, and FI_COLLEC‐
1146 TIVE.
1147 Many applications will be able to ignore this field and rely solely on
1149 the fi_info::caps field. Use of this field provides fine grained con‐
1150 trol over the receive capabilities associated with an endpoint. It is
1151 useful when handling scalable endpoints, with multiple receive con‐
1152 texts, for example, and allows configuring a specific receive context
1153 with fewer capabilities than that supported by the endpoint or other
1154 receive contexts.
1155 mode
1157 The operational mode bits of the context. The mode bits will be a sub‐
1158 set of those associated with the endpoint. See the MODE section of
1159 fi_getinfo(3) for details. A mode value of 0 will be ignored on input
1160 to fi_getinfo(3), with the mode value of the fi_info structure used in‐
1161 stead. On return from fi_getinfo(3), the mode will be set only to
1162 those constraints specific to receive operations.
1163 op_flags - Default receive operation flags
1165 Flags that control the operation of operations submitted against the
1166 context. Applicable flags are listed in the Operation Flags section.
1167 msg_order - Message Ordering
1169 For a description of message ordering, see the msg_order field in the
1170 Transmit Context Attribute section. Receive context message ordering
1171 defines the order in which received transport message headers are pro‐
1172 cessed when received by an endpoint. When ordering is set, it indi‐
1173 cates that message headers will be processed in order, based on how the
1174 transmit side has identified the messages. Typically, this means that
1175 messages will be handled in order based on a message level sequence
1176 number.
1177 The following ordering flags, as defined for transmit ordering, also
1179 apply to the processing of received operations: FI_ORDER_NONE, FI_OR‐
1180 DER_RAR, FI_ORDER_RAW, FI_ORDER_RAS, FI_ORDER_WAR, FI_ORDER_WAW, FI_OR‐
1181 DER_WAS, FI_ORDER_SAR, FI_ORDER_SAW, FI_ORDER_SAS, FI_ORDER_RMA_RAR,
1182 FI_ORDER_RMA_RAW, FI_ORDER_RMA_WAR, FI_ORDER_RMA_WAW, FI_ORDER_ATOM‐
1183 IC_RAR, FI_ORDER_ATOMIC_RAW, FI_ORDER_ATOMIC_WAR, and FI_ORDER_ATOM‐
1184 IC_WAW.
1185 comp_order - Completion Ordering
1187 For a description of completion ordering, see the comp_order field in
1188 the Transmit Context Attribute section.
1189 FI_ORDER_DATA
1191 When set, this bit indicates that received data is written into
1192 memory in order. Data ordering applies to memory accessed as
1193 part of a single operation and between operations if message or‐
1194 dering is guaranteed.
1195 FI_ORDER_NONE
1197 No ordering is defined for completed operations. Receive opera‐
1198 tions may complete in any order, regardless of their submission
1199 order.
1200 FI_ORDER_STRICT
1202 Receive operations complete in the order in which they are pro‐
1203 cessed by the receive context, based on the receive side msg_or‐
1204 der attribute.
1205 total_buffered_recv
1207 This field is supported for backwards compatibility purposes. It is a
1208 hint to the provider of the total available space that may be needed to
1209 buffer messages that are received for which there is no matching re‐
1210 ceive operation. The provider may adjust or ignore this value. The
1211 allocation of internal network buffering among received message is
1212 provider specific. For instance, a provider may limit the size of mes‐
1213 sages which can be buffered or the amount of buffering allocated to a
1214 single message.
1215 If receive side buffering is disabled (total_buffered_recv = 0) and a
1217 message is received by an endpoint, then the behavior is dependent on
1218 whether resource management has been enabled (FI_RM_ENABLED has be set
1219 or not). See the Resource Management section of fi_domain.3 for fur‐
1220 ther clarification. It is recommended that applications enable re‐
1221 source management if they anticipate receiving unexpected messages,
1222 rather than modifying this value.
1223 size
1225 The size of the receive context. The mapping of the size value to re‐
1226 sources is provider specific, but it is directly related to the number
1227 of command entries allocated for the endpoint. A smaller size value
1228 consumes fewer hardware and software resources, while a larger size al‐
1229 lows queuing more transmit requests.
1230 While the size attribute guides the size of underlying endpoint receive
1232 queue, there is not necessarily a one-to-one mapping between a receive
1233 operation and a queue entry. A single receive operation may consume
1234 multiple queue entries; for example, one per scatter-gather entry. Ad‐
1235 ditionally, the size field is intended to guide the allocation of the
1236 endpoint's receive context. Specifically, for connectionless end‐
1237 points, there may be lower-level queues use to track communication on a
1238 per peer basis. The sizes of any lower-level queues may only be sig‐
1239 nificantly smaller than the endpoint's receive size, in order to reduce
1240 resource utilization.
1241 iov_limit
1243 This is the maximum number of IO vectors (scatter-gather elements) that
1244 a single posted operating may reference.
1245
1247 A scalable endpoint is a communication portal that supports multiple
1248 transmit and receive contexts. Scalable endpoints are loosely modeled
1249 after the networking concept of transmit/receive side scaling, also
1250 known as multi-queue. Support for scalable endpoints is domain specif‐
1251 ic. Scalable endpoints may improve the performance of multi-threaded
1252 and parallel applications, by allowing threads to access independent
1253 transmit and receive queues. A scalable endpoint has a single trans‐
1254 port level address, which can reduce the memory requirements needed to
1255 store remote addressing data, versus using standard endpoints. Scal‐
1256 able endpoints cannot be used directly for communication operations,
1257 and require the application to explicitly create transmit and receive
1258 contexts as described below.
1259 fi_tx_context
1261 Transmit contexts are independent transmit queues. Ordering and syn‐
1262 chronization between contexts are not defined. Conceptually a transmit
1263 context behaves similar to a send-only endpoint. A transmit context
1264 may be configured with fewer capabilities than the base endpoint and
1265 with different attributes (such as ordering requirements and inject
1266 size) than other contexts associated with the same scalable endpoint.
1267 Each transmit context has its own completion queue. The number of
1268 transmit contexts associated with an endpoint is specified during end‐
1269 point creation.
1270 The fi_tx_context call is used to retrieve a specific context, identi‐
1272 fied by an index (see above for details on transmit context at‐
1273 tributes). Providers may dynamically allocate contexts when fi_tx_con‐
1274 text is called, or may statically create all contexts when fi_endpoint
1275 is invoked. By default, a transmit context inherits the properties of
1276 its associated endpoint. However, applications may request context
1277 specific attributes through the attr parameter. Support for per trans‐
1278 mit context attributes is provider specific and not guaranteed.
1279 Providers will return the actual attributes assigned to the context
1280 through the attr parameter, if provided.
1281 fi_rx_context
1283 Receive contexts are independent receive queues for receiving incoming
1284 data. Ordering and synchronization between contexts are not guaran‐
1285 teed. Conceptually a receive context behaves similar to a receive-only
1286 endpoint. A receive context may be configured with fewer capabilities
1287 than the base endpoint and with different attributes (such as ordering
1288 requirements and inject size) than other contexts associated with the
1289 same scalable endpoint. Each receive context has its own completion
1290 queue. The number of receive contexts associated with an endpoint is
1291 specified during endpoint creation.
1292 Receive contexts are often associated with steering flows, that specify
1294 which incoming packets targeting a scalable endpoint to process. How‐
1295 ever, receive contexts may be targeted directly by the initiator, if
1296 supported by the underlying protocol. Such contexts are referred to as
1297 'named'. Support for named contexts must be indicated by setting the
1298 caps FI_NAMED_RX_CTX capability when the corresponding endpoint is cre‐
1299 ated. Support for named receive contexts is coordinated with address
1300 vectors. See fi_av(3) and fi_rx_addr(3).
1301 The fi_rx_context call is used to retrieve a specific context, identi‐
1303 fied by an index (see above for details on receive context attributes).
1304 Providers may dynamically allocate contexts when fi_rx_context is
1305 called, or may statically create all contexts when fi_endpoint is in‐
1306 voked. By default, a receive context inherits the properties of its
1307 associated endpoint. However, applications may request context specif‐
1308 ic attributes through the attr parameter. Support for per receive con‐
1309 text attributes is provider specific and not guaranteed. Providers
1310 will return the actual attributes assigned to the context through the
1311 attr parameter, if provided.
1312
1314 Shared contexts are transmit and receive contexts explicitly shared
1315 among one or more endpoints. A shareable context allows an application
1316 to use a single dedicated provider resource among multiple transport
1317 addressable endpoints. This can greatly reduce the resources needed to
1318 manage communication over multiple endpoints by multiplexing transmit
1319 and/or receive processing, with the potential cost of serializing ac‐
1320 cess across multiple endpoints. Support for shareable contexts is do‐
1321 main specific.
1322 Conceptually, shareable transmit contexts are transmit queues that may
1324 be accessed by many endpoints. The use of a shared transmit context is
1325 mostly opaque to an application. Applications must allocate and bind
1326 shared transmit contexts to endpoints, but operations are posted di‐
1327 rectly to the endpoint. Shared transmit contexts are not associated
1328 with completion queues or counters. Completed operations are posted to
1329 the CQs bound to the endpoint. An endpoint may only be associated with
1330 a single shared transmit context.
1331 Unlike shared transmit contexts, applications interact directly with
1333 shared receive contexts. Users post receive buffers directly to a
1334 shared receive context, with the buffers usable by any endpoint bound
1335 to the shared receive context. Shared receive contexts are not associ‐
1336 ated with completion queues or counters. Completed receive operations
1337 are posted to the CQs bound to the endpoint. An endpoint may only be
1338 associated with a single receive context, and all connectionless end‐
1339 points associated with a shared receive context must also share the
1340 same address vector.
1341 Endpoints associated with a shared transmit context may use dedicated
1343 receive contexts, and vice-versa. Or an endpoint may use shared trans‐
1344 mit and receive contexts. And there is no requirement that the same
1345 group of endpoints sharing a context of one type also share the context
1346 of an alternate type. Furthermore, an endpoint may use a shared con‐
1347 text of one type, but a scalable set of contexts of the alternate type.
1348 fi_stx_context
1350 This call is used to open a shareable transmit context (see above for
1351 details on the transmit context attributes). Endpoints associated with
1352 a shared transmit context must use a subset of the transmit context's
1353 attributes. Note that this is the reverse of the requirement for
1354 transmit contexts for scalable endpoints.
1355 fi_srx_context
1357 This allocates a shareable receive context (see above for details on
1358 the receive context attributes). Endpoints associated with a shared
1359 receive context must use a subset of the receive context's attributes.
1360 Note that this is the reverse of the requirement for receive contexts
1361 for scalable endpoints.
1362
1364 The following feature and description should be considered experimen‐
1365 tal. Until the experimental tag is removed, the interfaces, semantics,
1366 and data structures associated with socket endpoints may change between
1367 library versions.
1368 This section applies to endpoints of type FI_EP_SOCK_STREAM and
1370 FI_EP_SOCK_DGRAM, commonly referred to as socket endpoints.
1371 Socket endpoints are defined with semantics that allow them to more
1373 easily be adopted by developers familiar with the UNIX socket API, or
1374 by middleware that exposes the socket API, while still taking advantage
1375 of high-performance hardware features.
1376 The key difference between socket endpoints and other active endpoints
1378 are socket endpoints use synchronous data transfers. Buffers passed
1379 into send and receive operations revert to the control of the applica‐
1380 tion upon returning from the function call. As a result, no data
1381 transfer completions are reported to the application, and socket end‐
1382 points are not associated with completion queues or counters.
1383 Socket endpoints support a subset of message operations: fi_send,
1385 fi_sendv, fi_sendmsg, fi_recv, fi_recvv, fi_recvmsg, and fi_inject.
1386 Because data transfers are synchronous, the return value from send and
1387 receive operations indicate the number of bytes transferred on success,
1388 or a negative value on error, including -FI_EAGAIN if the endpoint can‐
1389 not send or receive any data because of full or empty queues, respec‐
1390 tively.
1391 Socket endpoints are associated with event queues and address vectors,
1393 and process connection management events asynchronously, similar to
1394 other endpoints. Unlike UNIX sockets, socket endpoint must still be
1395 declared as either active or passive.
1396 Socket endpoints behave like non-blocking sockets. In order to support
1398 select and poll semantics, active socket endpoints are associated with
1399 a file descriptor that is signaled whenever the endpoint is ready to
1400 send and/or receive data. The file descriptor may be retrieved using
1401 fi_control.
1402
1404 Operation flags are obtained by OR-ing the following flags together.
1405 Operation flags define the default flags applied to an endpoint's data
1406 transfer operations, where a flags parameter is not available. Data
1407 transfer operations that take flags as input override the op_flags val‐
1408 ue of transmit or receive context attributes of an endpoint.
1409 FI_COMMIT_COMPLETE
1411 Indicates that a completion should not be generated (locally or
1412 at the peer) until the result of an operation have been made
1413 persistent. See fi_cq(3) for additional details on completion
1414 semantics.
1415 FI_COMPLETION
1417 Indicates that a completion queue entry should be written for
1418 data transfer operations. This flag only applies to operations
1419 issued on an endpoint that was bound to a completion queue with
1420 the FI_SELECTIVE_COMPLETION flag set, otherwise, it is ignored.
1421 See the fi_ep_bind section above for more detail.
1422 FI_DELIVERY_COMPLETE
1424 Indicates that a completion should be generated when the opera‐
1425 tion has been processed by the destination endpoint(s). See
1426 fi_cq(3) for additional details on completion semantics.
1427 FI_INJECT
1429 Indicates that all outbound data buffers should be returned to
1430 the user's control immediately after a data transfer call re‐
1431 turns, even if the operation is handled asynchronously. This
1432 may require that the provider copy the data into a local buffer
1433 and transfer out of that buffer. A provider can limit the total
1434 amount of send data that may be buffered and/or the size of a
1435 single send that can use this flag. This limit is indicated us‐
1436 ing inject_size (see inject_size above).
1437 FI_INJECT_COMPLETE
1439 Indicates that a completion should be generated when the source
1440 buffer(s) may be reused. See fi_cq(3) for additional details on
1441 completion semantics.
1442 FI_MULTICAST
1444 Indicates that data transfers will target multicast addresses by
1445 default. Any fi_addr_t passed into a data transfer operation
1446 will be treated as a multicast address.
1447 FI_MULTI_RECV
1449 Applies to posted receive operations. This flag allows the user
1450 to post a single buffer that will receive multiple incoming mes‐
1451 sages. Received messages will be packed into the receive buffer
1452 until the buffer has been consumed. Use of this flag may cause
1453 a single posted receive operation to generate multiple comple‐
1454 tions as messages are placed into the buffer. The placement of
1455 received data into the buffer may be subjected to provider spe‐
1456 cific alignment restrictions. The buffer will be released by
1457 the provider when the available buffer space falls below the
1458 specified minimum (see FI_OPT_MIN_MULTI_RECV).
1459 FI_TRANSMIT_COMPLETE
1461 Indicates that a completion should be generated when the trans‐
1462 mit operation has completed relative to the local provider. See
1463 fi_cq(3) for additional details on completion semantics.
1464
1466 Users should call fi_close to release all resources allocated to the
1467 fabric endpoint.
1468 Endpoints allocated with the FI_CONTEXT or FI_CONTEXT2 mode bits set
1470 must typically provide struct fi_context(2) as their per operation con‐
1471 text parameter. (See fi_getinfo.3 for details.) However, when FI_SE‐
1472 LECTIVE_COMPLETION is enabled to suppress CQ completion entries, and an
1473 operation is initiated without the FI_COMPLETION flag set, then the
1474 context parameter is ignored. An application does not need to pass in
1475 a valid struct fi_context(2) into such data transfers.
1476 Operations that complete in error that are not associated with valid
1478 operational context will use the endpoint context in any error report‐
1479 ing structures.
1480 Although applications typically associate individual completions with
1482 either completion queues or counters, an endpoint can be attached to
1483 both a counter and completion queue. When combined with using selec‐
1484 tive completions, this allows an application to use counters to track
1485 successful completions, with a CQ used to report errors. Operations
1486 that complete with an error increment the error counter and generate a
1487 CQ completion event.
1488 As mentioned in fi_getinfo(3), the ep_attr structure can be used to
1490 query providers that support various endpoint attributes. fi_getinfo
1491 can return provider info structures that can support the minimal set of
1492 requirements (such that the application maintains correctness). Howev‐
1493 er, it can also return provider info structures that exceed application
1494 requirements. As an example, consider an application requesting
1495 msg_order as FI_ORDER_NONE. The resulting output from fi_getinfo may
1496 have all the ordering bits set. The application can reset the ordering
1497 bits it does not require before creating the endpoint. The provider is
1498 free to implement a stricter ordering than is required by the applica‐
1499 tion.
1500
1502 Returns 0 on success. On error, a negative value corresponding to fab‐
1503 ric errno is returned. For fi_cancel, a return value of 0 indicates
1504 that the cancel request was submitted for processing.
1505 Fabric errno values are defined in rdma/fi_errno.h.
1507
1509 -FI_EDOMAIN
1510 A resource domain was not bound to the endpoint or an attempt
1511 was made to bind multiple domains.
1512 -FI_ENOCQ
1514 The endpoint has not been configured with necessary event queue.
1515 -FI_EOPBADSTATE
1517 The endpoint's state does not permit the requested operation.
1518
1520 fi_getinfo(3), fi_domain(3), fi_cq(3) fi_msg(3), fi_tagged(3),
1521 fi_rma(3)
1522
1524 OpenFabrics.
1525Libfabric Programmer's Manual 2021-01-19 fi_endpoint(3)