fi_getinfo(3) Libfabric Programmer's Manual
fi_getinfo / fi_freeinfo - Obtain / free fabric interface information
fi_allocinfo / fi_dupinfo - Allocate / duplicate an fi_info structure
- Interface version requested by application.
- Optional, name or fabric address to resolve.
- Optional, service name or port number of address.
- Operation flags for the fi_getinfo call.
- Reference to an fi_info structure that specifies criteria for selecting the returned fabric information.
- A pointer to a linked list of fi_info structures containing response information.
fi_getinfo returns information about available fabric services for reaching specified node or service, subject to any provided hints. Callers may specify NULL for node, service, and hints in order to retrieve information about what providers are available and their optimal usage models. If no matching fabric information is available, info will be set to NULL and the call will return -FI_ENODATA.
Based on the input hints, node, and service parameters, a list of fabric domains and endpoints will be returned. Each fi_info structure will describe an endpoint that meets the application’s specified communication criteria. Each endpoint will be associated with a domain. Applications can restrict the number of returned endpoints by including additional criteria in their search hints. Relaxing or eliminating input hints will increase the number and type of endpoints that are available. Providers that return multiple endpoints to a single fi_getinfo call should return the endpoints that are highest performing first. Providers may indicate that an endpoint and domain can support additional capabilities than those requested by the user only if such support will not adversely affect application performance or security.
The version parameter is used by the application to request the desired version of the interfaces. The version determines the format of all data structures used by any of the fabric interfaces. Applications should use the FI_VERSION(major, minor) macro to indicate the version, with hard-coded integer values for the major and minor values. The FI_MAJOR_VERSION and FI_MINOR_VERSION enum values defined in fabric.h specify the latest version of the installed library. However, it is recommended that the integer values for FI_MAJOR_VERSION and FI_MINOR_VERSION be used, rather than referencing the enum types in order to ensure compatibility with future versions of the library. This protects against the application being built from source against a newer version of the library that introduces new fields to data structures, which would not be initialized by the application.
Node, service, or hints may be provided, with any combination being supported. If node is provided, fi_getinfo will attempt to resolve the fabric address to the given node. If node is not given, fi_getinfo will attempt to resolve the fabric addressing information based on the provided hints.
The hints parameter, if provided, may be used to limit the resulting output as indicated below. As a general rule, specifying a non-zero value for input hints indicates that a provider must support the requested value or fail the operation with -FI_ENODATA. With the exception of mode bits, hints that are set to zero are treated as a wildcard. A zeroed hint value results in providers either returning a default value or a value that works best for their implementation. Mode bits that are set to zero indicate the application does not support any modes.
The caller must call fi_freeinfo to release fi_info structures returned by this call.
The fi_allocinfo call will allocate and zero an fi_info structure and all related substructures. The fi_dupinfo will duplicate a single fi_info structure and all the substructures within it.
- Pointer to the next fi_info structure in the list. Will be NULL if no more structures exist.
- caps - fabric interface capabilities
- If specified, indicates the desired capabilities of the fabric interfaces. Supported capabilities are listed in the Capabilities section below.
- Operational modes supported by the application. See the Mode section below.
- addr_format - address format
- If specified, indicates the format of addresses referenced by the fabric interfaces and data structures. Supported formats are listed in the Addressing formats section below.
- src_addrlen - source address length
- Indicates the length of the source address. This value must be > 0 if src_addr is non-NULL. This field will be ignored in hints if FI_SOURCE is specified, or src_addr is NULL.
- dest_addrlen - destination address length
- Indicates the length of the destination address. This value must be > 0 if dest_addr is non-NULL. This field will be ignored in hints unless the node and service parameters are NULL or FI_SOURCE is specified, or if dst_addr is NULL.
- src_addr - source address
- If specified, indicates the source address. This field will be ignored in hints if FI_SOURCE is specified. On output a provider shall return an address that corresponds to the indicated fabric or domain, with the format indicated by the returned addr_format field.
- dest_addr - destination address
- If specified, indicates the destination address. This field will be ignored in hints unless the node and service parameters are NULL or FI_SOURCE is specified.
- handle - provider context handle
- References a provider specific handle. The use of this field
is operation specific. Unless its use is described for a given operation,
the handle field must be NULL. It is commonly used by applications
that make use of connection-oriented endpoints. For other applications,
the field should usually be NULL.
This field is used when processing connection requests and responses. It is also used to inherit endpoint’s attributes. See fi_eq(3), fi_reject(3), and fi_endpoint(3) .
- tx_attr - transmit context attributes
- Optionally supplied transmit context attributes. Transmit context attributes may be specified and returned as part of fi_getinfo. When provided as hints, requested values of struct fi_tx_ctx_attr should be set. On output, the actual transmit context attributes that can be provided will be returned. Output values will be greater than or equal to the requested input values.
- rx_attr - receive context attributes
- Optionally supplied receive context attributes. Receive context attributes may be specified and returned as part of fi_getinfo. When provided as hints, requested values of struct fi_rx_ctx_attr should be set. On output, the actual receive context attributes that can be provided will be returned. Output values will be greater than or or equal to the requested input values.
- ep_attr - endpoint attributes
- Optionally supplied endpoint attributes. Endpoint attributes may be specified and returned as part of fi_getinfo. When provided as hints, requested values of struct fi_ep_attr should be set. On output, the actual endpoint attributes that can be provided will be returned. Output values will be greater than or equal to requested input values. See fi_endpoint(3) for details.
- domain_attr - domain attributes
- Optionally supplied domain attributes. Domain attributes may be specified and returned as part of fi_getinfo. When provided as hints, requested values of struct fi_domain_attr should be set. On output, the actual domain attributes that can be provided will be returned. Output values will be greater than or equal to requested input values. See fi_domain(3) for details.
- fabric_attr - fabric attributes
- Optionally supplied fabric attributes. Fabric attributes may be specified and returned as part of fi_getinfo. When provided as hints, requested values of struct fi_fabric_attr should be set. On output, the actual fabric attributes that can be provided will be returned. See fi_fabric(3) for details.
Interface capabilities are obtained by OR-ing the following flags together. If capabilities in the hint parameter are set to 0, the underlying provider will return the set of capabilities which are supported. Otherwise, providers will return data matching the specified set of capabilities. Providers may indicate support for additional capabilities beyond those requested when the use of expanded capabilities will not adversely affect performance or expose the application to communication beyond that which was requested. Applications may use this feature to request a minimal set of requirements, then check the returned capabilities to enable additional optimizations.
- Specifies that an endpoint should support sending and receiving
messages or datagrams. Message capabilities imply support for send
and/or receive queues. Endpoints supporting this capability support
operations defined by struct fi_ops_msg.
The caps may be used to specify or restrict the type of messaging operations that are supported. In the absence of any relevant flags, FI_MSG implies the ability to send and receive messages. Applications can use the FI_SEND and FI_RECV flags to optimize an endpoint as send-only or receive-only.
- Specifies that the endpoint should support RMA read and write operations. Endpoints supporting this capability support operations defined by struct fi_ops_rma. In the absence of any relevant flags, FI_RMA implies the ability to initiate and be the target of remote memory reads and writes. Applications can use the FI_READ, FI_WRITE, FI_REMOTE_READ, and FI_REMOTE_WRITE flags to restrict the types of RMA operations supported by an endpoint.
- Specifies that the endpoint should handle tagged message transfers. Tagged message transfers associate a user-specified key or tag with each message that is used for matching purposes at the remote side. Endpoints supporting this capability support operations defined by struct fi_ops_tagged. In the absence of any relevant flags, FI_TAGGED implies the ability to send and receive tagged messages. Applications can use the FI_SEND and FI_RECV flags to optimize an endpoint as send-only or receive-only.
- Specifies that the endpoint supports some set of atomic operations. Endpoints supporting this capability support operations defined by struct fi_ops_atomic. In the absence of any relevant flags, FI_ATOMIC implies the ability to initiate and be the target of remote atomic reads and writes. Applications can use the FI_READ, FI_WRITE, FI_REMOTE_READ, and FI_REMOTE_WRITE flags to restrict the types of atomic operations supported by an endpoint.
- Requests that endpoints which support multiple receive contexts allow an initiator to target (or name) a specific receive context as part of a data transfer operation.
- Requests that the communication endpoint use the source address of an incoming message when matching it with a receive buffer. If this capability is not set, then the src_addr parameter for msg and tagged receive operations is ignored.
- Specifies that the endpoint must support the FI_MULTI_RECV flag when posting receive buffers.
- Requests that the endpoint return source addressing data as part of its completion data. This capability only applies to connectionless endpoints. Note that returning source address information may require that the provider perform address translation and/or look-up based on data available in the underlying protocol in order to provide the requested data, which may adversely affect performance.
- Indicates that the user requires an endpoint capable of initiating reads against remote memory regions. This flag requires that FI_RMA and/or FI_ATOMIC be set.
- Indicates that the user requires an endpoint capable of initiating writes against remote memory regions. This flag requires that FI_RMA and/or FI_ATOMIC be set.
- Indicates that the user requires an endpoint capable of sending message data transfers. Message transfers include base message operations as well as tagged message functionality.
- Indicates that the user requires an endpoint capable of receiving message data transfers. Message transfers include base message operations as well as tagged message functionality.
- Indicates that the user requires an endpoint capable of receiving read memory operations from remote endpoints. This flag requires that FI_RMA and/or FI_ATOMIC be set.
- Indicates that the user requires an endpoint capable of receiving write memory operations from remote endpoints. This flag requires that FI_RMA and/or FI_ATOMIC be set.
- Requests that an endpoint support the generation of completion events when it is the target of an RMA and/or atomic operation. This flag requires that FI_REMOTE_READ and/or FI_REMOTE_WRITE be enabled on the endpoint.
- Indicates that the endpoint should support triggered operations. Endpoints support this capability must meet the usage model as described by fi_trigger.3.
- Indicates that the endpoint support the FI_FENCE flag on data transfer operations. Support requires tracking that all previous transmit requests to a specified remote endpoint complete prior to initiating the fenced operation. Fenced operations are often used to enforce ordering between operations that are not otherwise guaranteed by the underlying provider or protocol.
Capabilities may be grouped into two general categories: primary and secondary. Primary capabilities must explicitly be requested by an application, and a provider must enable support for only those primary capabilities which were selected. Secondary capabilities may optionally be requested by an application. If requested, a provider must support the capability or fail the fi_getinfo request (FI_ENODATA). A provider may optionally report non-selected secondary capabilities if doing so would not compromise performance or security.
Primary capabilities: FI_MSG, FI_RMA, FI_TAGGED, FI_ATOMIC, FI_NAMED_RX_CTX, FI_DIRECTED_RECV, FI_READ, FI_WRITE, FI_RECV, FI_SEND, FI_REMOTE_READ, and FI_REMOTE_WRITE.
Secondary capabilities: FI_MULTI_RECV, FI_SOURCE, FI_RMA_EVENT, FI_TRIGGER, FI_FENCE.
The operational mode bits are used to convey requirements that an application must adhere to when using the fabric interfaces. Modes specify optimal ways of accessing the reported endpoint or domain. Applications that are designed to support a specific mode of operation may see improved performance when that mode is desired by the provider. It is recommended that providers support applications that disable any provider preferred modes.
On input to fi_getinfo, applications set the mode bits that they support. On output, providers will clear mode bits that are not necessary to achieve high-performance. Mode bits that remain set indicate application requirements for using the fabric interfaces created using the returned fi_info. The set of modes are listed below. If a NULL hints structure is provided, then the provider’s supported set of modes will be returned in the info structure(s).
- Specifies that the provider requires that applications use struct fi_context as their per operation context parameter. This structure should be treated as opaque to the application. For performance reasons, this structure must be allocated by the user, but may be used by the fabric provider to track the operation. Typically, users embed struct fi_context within their own context structure. The struct fi_context must remain valid until the corresponding operation completes or is successfully canceled. As such, fi_context should NOT be allocated on the stack. Doing so is likely to result in stack corruption that will be difficult to debug. Users should not update or interpret the fields in this structure, or reuse it until the original operation has completed. The structure is specified in rdma/fabric.h.
- The provider is optimized around having applications register memory for locally accessed data buffers. Data buffers used in send and receive operations and as the source buffer for RMA and atomic operations must be registered by the application for access domains opened with this capability.
- Message prefix mode indicates that an application will provide
buffer space in front of all message send and receive buffers for
use by the provider. Typically, the provider uses this space to
implement a protocol, with the protocol headers being written into
the prefix area. The contents of the prefix space should be treated
as opaque. The use of FI_MSG_PREFIX may improve application
performance over certain providers by reducing the number of IO
vectors referenced by underlying hardware and eliminating provider
FI_MSG_PREFIX only applies to send and receive operations, including tagged sends and receives. RMA and atomics do not require the application to provide prefix buffers. Prefix buffer space must be provided with all sends and receives, regardless of the size of the transfer or other transfer options. The ownership of prefix buffers is treated the same as the corresponding message buffers, but the size of the prefix buffer is not counted toward any message limits, including inject.
Applications that support prefix mode must supply buffer space before their own message data. The size of space that must be provided is specified by the msg_prefix_size endpoint attribute. Providers are required to define a msg_prefix_size that is a multiple of 8 bytes. Additionally, applications may receive provider generated packets that do not contain application data. Such received messages will indicate a transfer size of that is equal to or smaller than msg_prefix_size.
The buffer pointer given to all send and receive operations must point to the start of the prefix region of the buffer (as opposed to the payload). For scatter-gather send/recv operations, the prefix buffer must be a contiguous region, though it may or may not be directly adjacent to the payload portion of the buffer.
- Applications can reference multiple data buffers as part of a single operation through the use of IO vectors (SGEs). Typically, the contents of an IO vector are copied by the provider into an internal buffer area, or directly to the underlying hardware. However, when a large number of IOV entries are supported, IOV buffering may have a negative impact on performance and memory consumption. The FI_ASYNC_IOV mode indicates that the application must provide the buffering needed for the IO vectors. When set, an application must not modify an IO vector, including any related memory descriptor array, until the associated operation has completed.
- This mode bit only applies to data transfers that set FI_REMOTE_CQ_DATA. When set, a data transfer that carries remote CQ data will consume a receive buffer at the target. This is true even for operations that would normally not consume posted receive buffers, such as RMA write operations.
Multiple fabric interfaces take as input either a source or destination address parameter. This includes struct fi_info (src_addr and dest_addr), CM calls (getname, getpeer, connect, join, and leave), and AV calls (insert, lookup, and straddr). The fi_info addr_format field indicates the expected address format for these operations.
A provider may support one or more of the following addressing
formats. In some cases, a selected addressing format may need to be
translated or mapped into an address which is native to the
- FI_FORMAT_UNSPEC indicates that a provider specific address format should be selected. Provider specific addresses may be protocol specific or a vendor proprietary format. Applications that select FI_FORMAT_UNSPEC should be prepared to treat returned addressing data as opaque. FI_FORMAT_UNSPEC targets apps which make use of an out of band address exchange. Applications which use FI_FORMAT_UNSPEC may use fi_getname() to obtain a provider specific address assigned to an allocated endpoint.
- Address is of type sockaddr. The specific socket address format will be determined at run time by interfaces examining the sa_family field.
- Address is of type sockaddr_in (IPv4).
- Address is of type sockaddr_in6 (IPv6).
- Address is of type sockaddr_ib (defined in Linux kernel source)
- Address is an Intel proprietary format that is used with their PSMX (extended performance scaled messaging) protocol.
The operation of the fi_getinfo call may be controlled through the use of input flags. Valid flags include the following.
- Indicates that the node parameter is a numeric string representation of a fabric address, such as a dotted decimal IP address. Use of this flag will suppress any lengthy name resolution protocol.
- Indicates that the node and service parameters specify the local source address to associate with an endpoint. If specified, either the node and/or service parameter must be non-NULL. This flag is often used with passive endpoints.
- Indicates that the caller is only querying for what providers are potentially available. All providers will return exactly one fi_info struct, regardless of whether that provider is usable on the current platform or not. The returned fi_info struct will contain default values for all members, with the exception of fabric_attr. The fabric_attr member will have the prov_name and prov_version values filled in.
fi_getinfo() returns 0 on success. On error, fi_getinfo() returns a
negative value corresponding to fabric errno. Fabric errno values are
fi_allocinfo() returns a pointer to a new fi_info structure on success, or NULL on error. fi_dupinfo() duplicates a single fi_info structure and all the substructures within it, returning a pointer to the new fi_info structure on success, or NULL on error. Both calls require that the returned fi_info structure be freed via fi_freeinfo().
- The specified endpoint or domain capability or operation flags are invalid.
- Indicates that there was insufficient memory to complete the operation.
- Indicates that no providers could be found which support the requested fabric information.
If hints are provided, the operation will be controlled by the values that are supplied in the various fields (see section on fi_info). Applications that require specific communication interfaces, domains, capabilities or other requirements, can specify them using fields in hints. Libfabric returns a linked list in info that points to a list of matching interfaces. info is set to NULL if there are no communication interfaces or none match the input hints.
If node is provided, fi_getinfo will attempt to resolve the fabric address to the given node. If node is not provided, fi_getinfo will attempt to resolve the fabric addressing information based on the provided hints. The caller must call fi_freeinfo to release fi_info structures returned by fi_getinfo.
If neither node, service or hints are provided, then fi_getinfo simply returns the list all available communication interfaces.
Multiple threads may call
fi_getinfo “simultaneously, without any requirement for serialization.”