1fi_domain(3) Libfabric v1.7.0 fi_domain(3)
2
6 fi_domain - Open a fabric access domain
7
9 #include <rdma/fabric.h>
10 #include <rdma/fi_domain.h>
12 int fi_domain(struct fid_fabric *fabric, struct fi_info *info,
14 struct fid_domain **domain, void *context);
15 int fi_close(struct fid *domain);
17 int fi_domain_bind(struct fid_domain *domain, struct fid *eq,
19 uint64_t flags);
20 int fi_open_ops(struct fid *domain, const char *name, uint64_t flags,
22 void **ops, void *context);
23
25 fabric Fabric domain
26 info Fabric information, including domain capabilities and at‐
28 tributes.
29 domain An opened access domain.
31 context
33 User specified context associated with the domain. This context
34 is returned as part of any asynchronous event associated with
35 the domain.
36 eq Event queue for asynchronous operations initiated on the domain.
38 name Name associated with an interface.
40 ops Fabric interface operations.
42
44 An access domain typically refers to a physical or virtual NIC or hard‐
45 ware port; however, a domain may span across multiple hardware compo‐
46 nents for fail-over or data striping purposes. A domain defines the
47 boundary for associating different resources together. Fabric re‐
48 sources belonging to the same domain may share resources.
49 fi_domain
51 Opens a fabric access domain, also referred to as a resource domain.
52 Fabric domains are identified by a name. The properties of the opened
53 domain are specified using the info parameter.
54 fi_open_ops
56 fi_open_ops is used to open provider specific interfaces. Provider in‐
57 terfaces may be used to access low-level resources and operations that
58 are specific to the opened resource domain. The details of domain in‐
59 terfaces are outside the scope of this documentation.
60 fi_domain_bind
62 Associates an event queue with the domain. An event queue bound to a
63 domain will be the default EQ associated with asynchronous control
64 events that occur on the domain or active endpoints allocated on a do‐
65 main. This includes CM events. Endpoints may direct their control
66 events to alternate EQs by binding directly with the EQ.
67 Binding an event queue to a domain with the FI_REG_MR flag indicates
69 that the provider should perform all memory registration operations
70 asynchronously, with the completion reported through the event queue.
71 If an event queue is not bound to the domain with the FI_REG_MR flag,
72 then memory registration requests complete synchronously.
73 See fi_av_bind(3), fi_ep_bind(3), fi_mr_bind(3), fi_pep_bind(3), and
75 fi_scalable_ep_bind(3) for more information.
76 fi_close
78 The fi_close call is used to release all resources associated with a
79 domain or interface. All objects associated with the opened domain
80 must be released prior to calling fi_close, otherwise the call will re‐
81 turn -FI_EBUSY.
82
84 The fi_domain_attr structure defines the set of attributes associated
85 with a domain.
86 struct fi_domain_attr {
88 struct fid_domain *domain;
89 char *name;
90 enum fi_threading threading;
91 enum fi_progress control_progress;
92 enum fi_progress data_progress;
93 enum fi_resource_mgmt resource_mgmt;
94 enum fi_av_type av_type;
95 int mr_mode;
96 size_t mr_key_size;
97 size_t cq_data_size;
98 size_t cq_cnt;
99 size_t ep_cnt;
100 size_t tx_ctx_cnt;
101 size_t rx_ctx_cnt;
102 size_t max_ep_tx_ctx;
103 size_t max_ep_rx_ctx;
104 size_t max_ep_stx_ctx;
105 size_t max_ep_srx_ctx;
106 size_t cntr_cnt;
107 size_t mr_iov_limit;
108 uint64_t caps;
109 uint64_t mode;
110 uint8_t *auth_key;
111 size_t auth_key_size;
112 size_t max_err_data;
113 size_t mr_cnt;
114 };
115 domain
117 On input to fi_getinfo, a user may set this to an opened domain in‐
118 stance to restrict output to the given domain. On output from fi_get‐
119 info, if no domain was specified, but the user has an opened instance
120 of the named domain, this will reference the first opened instance. If
121 no instance has been opened, this field will be NULL.
122 Name
124 The name of the access domain.
125 Multi-threading Support (threading)
127 The threading model specifies the level of serialization required of an
128 application when using the libfabric data transfer interfaces. Control
129 interfaces are always considered thread safe, and may be accessed by
130 multiple threads. Applications which can guarantee serialization in
131 their access of provider allocated resources and interfaces enables a
132 provider to eliminate lower-level locks.
133 FI_THREAD_UNSPEC
135 This value indicates that no threading model has been defined.
136 It may be used on input hints to the fi_getinfo call. When
137 specified, providers will return a threading model that allows
138 for the greatest level of parallelism.
139 FI_THREAD_SAFE
141 A thread safe serialization model allows a multi-threaded appli‐
142 cation to access any allocated resources through any interface
143 without restriction. All providers are required to support
144 FI_THREAD_SAFE.
145 FI_THREAD_FID
147 A fabric descriptor (FID) serialization model requires applica‐
148 tions to serialize access to individual fabric resources associ‐
149 ated with data transfer operations and completions. Multiple
150 threads must be serialized when accessing the same endpoint,
151 transmit context, receive context, completion queue, counter,
152 wait set, or poll set. Serialization is required only by
153 threads accessing the same object.
154 For example, one thread may be initiating a data transfer on an end‐
156 point, while another thread reads from a completion queue associated
157 with the endpoint.
158 Serialization to endpoint access is only required when accessing the
160 same endpoint data flow. Multiple threads may initiate transfers on
161 different transmit contexts of the same endpoint without serializing,
162 and no serialization is required between the submission of data trans‐
163 mit requests and data receive operations.
164 In general, FI_THREAD_FID allows the provider to be implemented without
166 needing internal locking when handling data transfers. Conceptually,
167 FI_THREAD_FID maps well to providers that implement fabric services in
168 hardware and provide separate command queues to different data flows.
169 FI_THREAD_ENDPOINT
171 The endpoint threading model is similar to FI_THREAD_FID, but
172 with the added restriction that serialization is required when
173 accessing the same endpoint, even if multiple transmit and re‐
174 ceive contexts are used. Conceptually, FI_THREAD_ENDPOINT maps
175 well to providers that implement fabric services in hardware but
176 use a single command queue to access different data flows.
177 FI_THREAD_COMPLETION
179 The completion threading model is intended for providers that
180 make use of manual progress. Applications must serialize access
181 to all objects that are associated through the use of having a
182 shared completion structure. This includes endpoint, transmit
183 context, receive context, completion queue, counter, wait set,
184 and poll set objects.
185 For example, threads must serialize access to an endpoint and its bound
187 completion queue(s) and/or counters. Access to endpoints that share
188 the same completion queue must also be serialized.
189 The use of FI_THREAD_COMPLETION can increase parallelism over
191 FI_THREAD_SAFE, but requires the use of isolated resources.
192 FI_THREAD_DOMAIN
194 A domain serialization model requires applications to serialize
195 access to all objects belonging to a domain.
196 Progress Models (control_progress / data_progress)
198 Progress is the ability of the underlying implementation to complete
199 processing of an asynchronous request. In many cases, the processing
200 of an asynchronous request requires the use of the host processor. For
201 example, a received message may need to be matched with the correct
202 buffer, or a timed out request may need to be retransmitted. For per‐
203 formance reasons, it may be undesirable for the provider to allocate a
204 thread for this purpose, which will compete with the application
205 threads.
206 Control progress indicates the method that the provider uses to make
208 progress on asynchronous control operations. Control operations are
209 functions which do not directly involve the transfer of application da‐
210 ta between endpoints. They include address vector, memory registra‐
211 tion, and connection management routines.
212 Data progress indicates the method that the provider uses to make
214 progress on data transfer operations. This includes message queue,
215 RMA, tagged messaging, and atomic operations, along with their comple‐
216 tion processing.
217 Progress frequently requires action being taken at both the transmit‐
219 ting and receiving sides of an operation. This is often a requirement
220 for reliable transfers, as a result of retry and acknowledgement pro‐
221 cessing.
222 To balance between performance and ease of use, two progress models are
224 defined.
225 FI_PROGRESS_UNSPEC
227 This value indicates that no progress model has been defined.
228 It may be used on input hints to the fi_getinfo call.
229 FI_PROGRESS_AUTO
231 This progress model indicates that the provider will make for‐
232 ward progress on an asynchronous operation without further in‐
233 tervention by the application. When FI_PROGRESS_AUTO is provid‐
234 ed as output to fi_getinfo in the absence of any progress hints,
235 it often indicates that the desired functionality is implemented
236 by the provider hardware or is a standard service of the operat‐
237 ing system.
238 All providers are required to support FI_PROGRESS_AUTO. However, if a
240 provider does not natively support automatic progress, forcing the use
241 of FI_PROGRESS_AUTO may result in threads being allocated below the
242 fabric interfaces.
243 FI_PROGRESS_MANUAL
245 This progress model indicates that the provider requires the use
246 of an application thread to complete an asynchronous request.
247 When manual progress is set, the provider will attempt to ad‐
248 vance an asynchronous operation forward when the application at‐
249 tempts to wait on or read an event queue, completion queue, or
250 counter where the completed operation will be reported.
251 Progress also occurs when the application processes a poll or
252 wait set that has been associated with the event or completion
253 queue.
254 Only wait operations defined by the fabric interface will result in an
256 operation progressing. Operating system or external wait functions,
257 such as select, poll, or pthread routines, cannot.
258 Manual progress requirements not only apply to endpoints that initiate
260 transmit operations, but also to endpoints that may be the target of
261 such operations. This holds true even if the target endpoint will not
262 generate completion events for the operations. For example, an end‐
263 point that acts purely as the target of RMA or atomic operations that
264 uses manual progress may still need application assistance to process
265 received operations.
266 Resource Management (resource_mgmt)
268 Resource management (RM) is provider and protocol support to protect
269 against overrunning local and remote resources. This includes local
270 and remote transmit contexts, receive contexts, completion queues, and
271 source and target data buffers.
272 When enabled, applications are given some level of protection against
274 overrunning provider queues and local and remote data buffers. Such
275 support may be built directly into the hardware and/or network proto‐
276 col, but may also require that checks be enabled in the provider soft‐
277 ware. By disabling resource management, an application assumes all re‐
278 sponsibility for preventing queue and buffer overruns, but doing so may
279 allow a provider to eliminate internal synchronization calls, such as
280 atomic variables or locks.
281 It should be noted that even if resource management is disabled, the
283 provider implementation and protocol may still provide some level of
284 protection against overruns. However, such protection is not guaran‐
285 teed. The following values for resource management are defined.
286 FI_RM_UNSPEC
288 This value indicates that no resource management model has been
289 defined. It may be used on input hints to the fi_getinfo call.
290 FI_RM_DISABLED
292 The provider is free to select an implementation and protocol
293 that does not protect against resource overruns. The applica‐
294 tion is responsible for resource protection.
295 FI_RM_ENABLED
297 Resource management is enabled for this provider domain.
298 The behavior of the various resource management options depends on
300 whether the endpoint is reliable or unreliable, as well as provider and
301 protocol specific implementation details, as shown in the following ta‐
302 ble. The table assumes that all peers enable or disable RM the same.
303 Resource DGRAM EP-no RM DGRAM EP-with RM RDM/MSG EP-no RDM/MSG EP-with
305 RM RM
306 ───────────────────────────────────────────────────────────────────────────────────
307 Tx Ctx undefined error EAGAIN undefined error EAGAIN
308 Rx Ctx undefined error EAGAIN undefined error EAGAIN
309 Tx CQ undefined error EAGAIN undefined error EAGAIN
310 Rx CQ undefined error EAGAIN undefined error EAGAIN
311 Target dropped dropped transmit error retried
312 EP
313 No Rx dropped dropped transmit error retried
314 Buffer
315 Rx Buf truncate or drop truncate or drop truncate or er‐ truncate or er‐
316 Overrun ror ror
317 Un‐ not applicable not applicable transmit error transmit error
318 matched
319 RMA
320 RMA not applicable not applicable transmit error transmit error
321 Overrun
322 The resource column indicates the resource being accessed by a data
324 transfer operation.
325 Tx Ctx / Rx Ctx
327 Refers to the transmit/receive contexts when a data transfer op‐
328 eration is submitted. When RM is enabled, attempting to submit
329 a request will fail if the context is full. If RM is disabled,
330 an undefined error (provider specific) will occur. Such errors
331 should be considered fatal to the context, and applications must
332 take steps to avoid queue overruns.
333 Tx CQ / Rx CQ
335 Refers to the completion queue associated with the Tx or Rx con‐
336 text when a local operation completes. When RM is disabled, ap‐
337 plications must take care to ensure that completion queues do
338 not get overrun. When an overrun occurs, an undefined, but fa‐
339 tal, error will occur affecting all endpoints associated with
340 the CQ. Overruns can be avoided by sizing the CQs appropriately
341 or by deferring the posting of a data transfer operation unless
342 CQ space is available to store its completion. When RM is en‐
343 abled, providers may use different mechanisms to prevent CQ
344 overruns. This includes failing (returning -FI_EAGAIN) the
345 posting of operations that could result in CQ overruns, or in‐
346 ternally retrying requests (which will be hidden from the appli‐
347 cation). See notes at the end of this section regarding CQ re‐
348 source management restrictions.
349 Target EP / No Rx Buffer
351 Target EP refers to resources associated with the endpoint that
352 is the target of a transmit operation. This includes the target
353 endpoint's receive queue, posted receive buffers (no Rx buf‐
354 fers), the receive side completion queue, and other related
355 packet processing queues. The defined behavior is that seen by
356 the initiator of a request. For FI_EP_DGRAM endpoints, if the
357 target EP queues are unable to accept incoming messages, re‐
358 ceived messages will be dropped. For reliable endpoints, if RM
359 is disabled, the transmit operation will complete in error. If
360 RM is enabled, the provider will internally retry the operation.
361 Rx Buffer Overrun
363 This refers to buffers posted to receive incoming tagged or un‐
364 tagged messages, with the behavior defined from the viewpoint of
365 the sender. The behavior for handling received messages that
366 are larger than the buffers provided by the application is
367 provider specific. Providers may either truncate the message
368 and report a successful completion, or fail the operation. For
369 datagram endpoints, failed sends will result in the message be‐
370 ing dropped. For reliable endpoints, send operations may com‐
371 plete successfully, yet be truncated at the receive side. This
372 can occur when the target side buffers received data until an
373 application buffer is made available. The completion status may
374 also be dependent upon the completion model selected byt the ap‐
375 plication (e.g. FI_DELIVERY_COMPLETE versus FI_TRANSMIT_COM‐
376 PLETE).
377 Unmatched RMA / RMA Overrun
379 Unmatched RMA and RMA overruns deal with the processing of RMA
380 and atomic operations. Unlike send operations, RMA operations
381 that attempt to access a memory address that is either not reg‐
382 istered for such operations, or attempt to access outside of the
383 target memory region will fail, resulting in a transmit error.
384 When a resource management error occurs on an endpoint, the endpoint is
386 transitioned into a disabled state. Any operations which have not al‐
387 ready completed will fail and be discarded. For unconnected endpoints,
388 the endpoint must be re-enabled before it will accept new data transfer
389 operations. For connected endpoints, the connection is torn down and
390 must be re-established.
391 There is one notable restriction on the protections offered by resource
393 management. This occurs when resource management is enabled on an end‐
394 point that has been bound to completion queue(s) using the FI_SELEC‐
395 TIVE_COMPLETION flag. Operations posted to such an endpoint may speci‐
396 fy that a successful completion should not generate a entry on the cor‐
397 responding completion queue. (I.e. the operation leaves the FI_COM‐
398 PLETION flag unset). In such situations, the provider is not required
399 to reserve an entry in the completion queue to handle the case where
400 the operation fails and does generate a CQ entry, which would effec‐
401 tively require tracking the operation to completion. Applications con‐
402 cerned with avoiding CQ overruns in the occurrence of errors must en‐
403 sure that there is sufficient space in the CQ to report failed opera‐
404 tions. This can typically be achieved by sizing the CQ to at least the
405 same size as the endpoint queue(s) that are bound to it.
406 AV Type (av_type)
408 Specifies the type of address vectors that are usable with this domain.
409 For additional details on AV type, see fi_av(3). The following values
410 may be specified.
411 FI_AV_UNSPEC
413 Any address vector format is requested and supported.
414 FI_AV_MAP
416 Only address vectors of type AV map are requested or supported.
417 FI_AV_TABLE
419 Only address vectors of type AV index are requested or support‐
420 ed.
421 Address vectors are only used by connectionless endpoints. Applica‐
423 tions that require the use of a specific type of address vector should
424 set the domain attribute av_type to the necessary value when calling
425 fi_getinfo. The value FI_AV_UNSPEC may be used to indicate that the
426 provider can support either address vector format. In this case, a
427 provider may return FI_AV_UNSPEC to indicate that either format is sup‐
428 portable, or may return another AV type to indicate the optimal AV type
429 supported by this domain.
430 Memory Registration Mode (mr_mode)
432 Defines memory registration specific mode bits used with this domain.
433 Full details on MR mode options are available in fi_mr(3). The follow‐
434 ing values may be specified.
435 FI_MR_LOCAL
437 The provider is optimized around having applications register
438 memory for locally accessed data buffers. Data buffers used in
439 send and receive operations and as the source buffer for RMA and
440 atomic operations must be registered by the application for ac‐
441 cess domains opened with this capability.
442 FI_MR_RAW
444 The provider requires additional setup as part of their memory
445 registration process. This mode is required by providers that
446 use a memory key that is larger than 64-bits.
447 FI_MR_VIRT_ADDR
449 Registered memory regions are referenced by peers using the vir‐
450 tual address of the registered memory region, rather than a
451 0-based offset.
452 FI_MR_ALLOCATED
454 Indicates that memory registration occurs on allocated data buf‐
455 fers, and physical pages must back all virtual addresses being
456 registered.
457 FI_MR_PROV_KEY
459 Memory registration keys are selected and returned by the
460 provider.
461 FI_MR_MMU_NOTIFY
463 Indicates that the application is responsible for notifying the
464 provider when the page tables referencing a registered memory
465 region may have been updated.
466 FI_MR_RMA_EVENT
468 Indicates that the memory regions associated with completion
469 counters must be explicitly enabled after being bound to any
470 counter.
471 FI_MR_ENDPOINT
473 Memory registration occurs at the endpoint level, rather than
474 domain.
475 FI_MR_UNSPEC
477 Defined for compatibility -- library versions 1.4 and earlier.
478 Setting mr_mode to 0 indicates that FI_MR_BASIC or FI_MR_SCAL‐
479 ABLE are requested and supported.
480 FI_MR_BASIC
482 Defined for compatibility -- library versions 1.4 and earlier.
483 Only basic memory registration operations are requested or sup‐
484 ported. This mode is equivalent to the FI_MR_VIRT_ADDR,
485 FI_MR_ALLOCATED, and FI_MR_PROV_KEY flags being set in later li‐
486 brary versions. This flag may not be used in conjunction with
487 other mr_mode bits.
488 FI_MR_SCALABLE
490 Defined for compatibility -- library versions 1.4 and earlier.
491 Only scalable memory registration operations are requested or
492 supported. Scalable registration uses offset based addressing,
493 with application selectable memory keys. For library versions
494 1.5 and later, this is the default if no mr_mode bits are set.
495 This flag may not be used in conjunction with other mr_mode
496 bits.
497 Buffers used in data transfer operations may require notifying the
499 provider of their use before a data transfer can occur. The mr_mode
500 field indicates the type of memory registration that is required, and
501 when registration is necessary. Applications that require the use of a
502 specific registration mode should set the domain attribute mr_mode to
503 the necessary value when calling fi_getinfo. The value FI_MR_UNSPEC
504 may be used to indicate support for any registration mode.
505 MR Key Size (mr_key_size)
507 Size of the memory region remote access key, in bytes. Applications
508 that request their own MR key must select a value within the range
509 specified by this value. Key sizes larger than 8 bytes require using
510 the FI_RAW_KEY mode bit.
511 CQ Data Size (cq_data_size)
513 Applications may include a small message with a data transfer that is
514 placed directly into a remote completion queue as part of a completion
515 event. This is referred to as remote CQ data (sometimes referred to as
516 immediate data). This field indicates the number of bytes that the
517 provider supports for remote CQ data. If supported (non-zero value is
518 returned), the minimum size of remote CQ data must be at least 4-bytes.
519 Completion Queue Count (cq_cnt)
521 The optimal number of completion queues supported by the domain, rela‐
522 tive to any specified or default CQ attributes. The cq_cnt value may
523 be a fixed value of the maximum number of CQs supported by the underly‐
524 ing hardware, or may be a dynamic value, based on the default at‐
525 tributes of an allocated CQ, such as the CQ size and data format.
526 Endpoint Count (ep_cnt)
528 The total number of endpoints supported by the domain, relative to any
529 specified or default endpoint attributes. The ep_cnt value may be a
530 fixed value of the maximum number of endpoints supported by the under‐
531 lying hardware, or may be a dynamic value, based on the default at‐
532 tributes of an allocated endpoint, such as the endpoint capabilities
533 and size. The endpoint count is the number of addressable endpoints
534 supported by the provider.
535 Transmit Context Count (tx_ctx_cnt)
537 The number of outbound command queues optimally supported by the
538 provider. For a low-level provider, this represents the number of com‐
539 mand queues to the hardware and/or the number of parallel transmit en‐
540 gines effectively supported by the hardware and caches. Applications
541 which allocate more transmit contexts than this value will end up shar‐
542 ing underlying resources. By default, there is a single transmit con‐
543 text associated with each endpoint, but in an advanced usage model, an
544 endpoint may be configured with multiple transmit contexts.
545 Receive Context Count (rx_ctx_cnt)
547 The number of inbound processing queues optimally supported by the
548 provider. For a low-level provider, this represents the number hard‐
549 ware queues that can be effectively utilized for processing incoming
550 packets. Applications which allocate more receive contexts than this
551 value will end up sharing underlying resources. By default, a single
552 receive context is associated with each endpoint, but in an advanced
553 usage model, an endpoint may be configured with multiple receive con‐
554 texts.
555 Maximum Endpoint Transmit Context (max_ep_tx_ctx)
557 The maximum number of transmit contexts that may be associated with an
558 endpoint.
559 Maximum Endpoint Receive Context (max_ep_rx_ctx)
561 The maximum number of receive contexts that may be associated with an
562 endpoint.
563 Maximum Sharing of Transmit Context (max_ep_stx_ctx)
565 The maximum number of endpoints that may be associated with a shared
566 transmit context.
567 Maximum Sharing of Receive Context (max_ep_srx_ctx)
569 The maximum number of endpoints that may be associated with a shared
570 receive context.
571 Counter Count (cntr_cnt)
573 The optimal number of completion counters supported by the domain. The
574 cq_cnt value may be a fixed value of the maximum number of counters
575 supported by the underlying hardware, or may be a dynamic value, based
576 on the default attributes of the domain.
577 MR IOV Limit (mr_iov_limit)
579 This is the maximum number of IO vectors (scatter-gather elements) that
580 a single memory registration operation may reference.
581 Capabilities (caps)
583 Domain level capabilities. Domain capabilities indicate domain level
584 features that are supported by the provider.
585 FI_LOCAL_COMM
587 At a conceptual level, this field indicates that the underlying
588 device supports loopback communication. More specifically, this
589 field indicates that an endpoint may communicate with other end‐
590 points that are allocated from the same underlying named domain.
591 If this field is not set, an application may need to use an al‐
592 ternate domain or mechanism (e.g. shared memory) to communicate
593 with peers that execute on the same node.
594 FI_REMOTE_COMM
596 This field indicates that the underlying provider supports com‐
597 munication with nodes that are reachable over the network. If
598 this field is not set, then the provider only supports communi‐
599 cation between processes that execute on the same node -- a
600 shared memory provider, for example.
601 FI_SHARED_AV
603 Indicates that the domain supports the ability to share address
604 vectors among multiple processes using the named address vector
605 feature.
606 See fi_getinfo(3) for a discussion on primary versus secondary capabil‐
608 ities. All domain capabilities are considered secondary capabilities.
609 mode
611 The operational mode bit related to using the domain.
612 FI_RESTRICTED_COMP
614 This bit indicates that the domain limits completion queues and
615 counters to only be used with endpoints, transmit contexts, and
616 receive contexts that have the same set of capability flags.
617 Default authorization key (auth_key)
619 The default authorization key to associate with endpoint and memory
620 registrations created within the domain. This field is ignored unless
621 the fabric is opened with API version 1.5 or greater.
622 Default authorization key length (auth_key_size)
624 The length in bytes of the default authorization key for the domain.
625 If set to 0, then no authorization key will be associated with end‐
626 points and memory registrations created within the domain unless speci‐
627 fied in the endpoint or memory registration attributes. This field is
628 ignored unless the fabric is opened with API version 1.5 or greater.
629 Max Error Data Size (max_err_data)
631 : The maximum amount of error data, in bytes, that may be returned as
632 part of a completion or event queue error. This value corresponds to
633 the err_data_size field in struct fi_cq_err_entry and struct
634 fi_eq_err_entry.
635 Memory Regions Count (mr_cnt)
637 The optimal number of memory regions supported by the domain, or end‐
638 point if the mr_mode FI_MR_ENDPOINT bit has been set. The mr_cnt value
639 may be a fixed value of the maximum number of MRs supported by the un‐
640 derlying hardware, or may be a dynamic value, based on the default at‐
641 tributes of the domain, such as the supported memory registration
642 modes. Applications can set the mr_cnt on input to fi_getinfo, in or‐
643 der to indicate their memory registration requirements. Doing so may
644 allow the provider to optimize any memory registration cache or lookup
645 tables.
646
648 Returns 0 on success. On error, a negative value corresponding to fab‐
649 ric errno is returned. Fabric errno values are defined in rdma/fi_er‐
650 rno.h.
651
653 Users should call fi_close to release all resources allocated to the
654 fabric domain.
655 The following fabric resources are associated with domains: active end‐
657 points, memory regions, completion event queues, and address vectors.
658 Domain attributes reflect the limitations and capabilities of the un‐
660 derlying hardware and/or software provider. They do not reflect system
661 limitations, such as the number of physical pages that an application
662 may pin or number of file descriptors that the application may open.
663 As a result, the reported maximums may not be achievable, even on a
664 lightly loaded systems, without an administrator configuring system re‐
665 sources appropriately for the installed provider(s).
666
668 fi_getinfo(3), fi_endpoint(3), fi_av(3), fi_ep(3), fi_eq(3), fi_mr(3)
669
671 OpenFabrics.
672Libfabric Programmer's Manual 2018-10-05 fi_domain(3)