fi_poll_open

fi_poll_open creates a new polling set. A poll set enables an optimized method for progressing asynchronous operations across multiple completion queues and counters and checking for their completions.

A poll set is defined with the following attributes.

struct fi_poll_attr {
    uint64_t             flags;     /* operation flags */
};

flags

Flags that set the default operation of the poll set. The use of this field is reserved and must be set to 0 by the caller.

fi_close

The fi_close call releases all resources associated with a poll set. The poll set must not be associated with any other resources prior to being closed, otherwise the call will return -FI_EBUSY.

fi_poll_add

Associates a completion queue or counter with a poll set.

fi_poll_del

Removes a completion queue or counter from a poll set.

fi_poll

Progresses all completion queues and counters associated with a poll set and checks for events. If events might have occurred, contexts associated with the completion queues and/or counters are returned. Completion queues will return their context if they are not empty. The context associated with a counter will be returned if the counter's success value or error value have changed since the last time fi_poll, fi_cntr_set, or fi_cntr_add were called. The number of contexts is limited to the size of the context array, indicated by the count parameter.

Note that fi_poll only indicates that events might be available. In some cases, providers may consume such events internally, to drive progress, for example. This can result in fi_poll returning false positives. Applications should drive their progress based on the results of reading events from a completion queue or reading counter values. The fi_poll function will always return all completion queues and counters that do have new events.

fi_wait_open

fi_wait_open allocates a new wait set. A wait set enables an optimized method of waiting for events across multiple completion queues and counters. Where possible, a wait set uses a single underlying wait object that is signaled when a specified condition occurs on an associated completion queue or counter.

The properties and behavior of a wait set are defined by struct fi_wait_attr.

struct fi_wait_attr {
    enum fi_wait_obj     wait_obj;  /* requested wait object */
    uint64_t             flags;     /* operation flags */
};

wait_obj

Wait sets are associated with specific wait object(s). Wait objects allow applications to block until the wait object is signaled, indicating that an event is available to be read. The following values may be used to specify the type of wait object associated with a wait set: FI_WAIT_UNSPEC, FI_WAIT_FD, FI_WAIT_MUTEX_COND, and FI_WAIT_YIELD.

- FI_WAIT_UNSPEC

Specifies that the user will only wait on the wait set using fabric interface calls, such as fi_wait. In this case, the underlying provider may select the most appropriate or highest performing wait object available, including custom wait mechanisms. Applications that select FI_WAIT_UNSPEC are not guaranteed to retrieve the underlying wait object.

- FI_WAIT_FD

Indicates that the wait set should use a single file descriptor as its wait mechanism, as exposed to the application. Internally, this may require the use of epoll in order to support waiting on a single file descriptor. File descriptor wait objects must be usable in the POSIX select(2) and poll(2), and Linux epoll(7) routines (if available). Provider signal an FD wait object by marking it as readable or with an error.

- FI_WAIT_MUTEX_COND

Specifies that the wait set should use a pthread mutex and cond variable as a wait object.

- FI_WAIT_POLLFD

This option is similar to FI_WAIT_FD, but allows the wait mechanism to use multiple file descriptors as its wait mechanism, as viewed by the application. The use of FI_WAIT_POLLFD can eliminate the need to use epoll to abstract away needing to check multiple file descriptors when waiting for events. The file descriptors must be usable in the POSIX select(2) and poll(2) routines, and match directly to being used with poll. See the NOTES section below for details on using pollfd.

- FI_WAIT_YIELD

Indicates that the wait set will wait without a wait object but instead yield on every wait.

flags

Flags that set the default operation of the wait set. The use of this field is reserved and must be set to 0 by the caller.

fi_close

The fi_close call releases all resources associated with a wait set. The wait set must not be bound to any other opened resources prior to being closed, otherwise the call will return -FI_EBUSY.

fi_wait

Waits on a wait set until one or more of its underlying wait objects is signaled.

fi_trywait

The fi_trywait call was introduced in libfabric version 1.3. The behavior of using native wait objects without the use of fi_trywait is provider specific and should be considered non-deterministic.

The fi_trywait() call is used in conjunction with native operating system calls to block on wait objects, such as file descriptors. The application must call fi_trywait and obtain a return value of FI_SUCCESS prior to blocking on a native wait object. Failure to do so may result in the wait object not being signaled, and the application not observing the desired events. The following pseudo-code demonstrates the use of fi_trywait in conjunction with the OS select(2) call.

fi_control(&cq->fid, FI_GETWAIT, (void *) &fd);
FD_ZERO(&fds);
FD_SET(fd, &fds);
while (1) {
    if (fi_trywait(&cq, 1) == FI_SUCCESS)
        select(fd + 1, &fds, NULL, &fds, &timeout);
    do {
        ret = fi_cq_read(cq, &comp, 1);
    } while (ret > 0);
}

fi_trywait() will return FI_SUCCESS if it is safe to block on the wait object(s) corresponding to the fabric descriptor(s), or -FI_EAGAIN if there are events queued on the fabric descriptor or if blocking could hang the application.

The call takes an array of fabric descriptors. For each wait object that will be passed to the native wait routine, the corresponding fabric descriptor should first be passed to fi_trywait. All fabric descriptors passed into a single fi_trywait call must make use of the same underlying wait object type.

The following types of fabric descriptors may be passed into fi_trywait: event queues, completion queues, counters, and wait sets. Applications that wish to use native wait calls should select specific wait objects when allocating such resources. For example, by setting the item's creation attribute wait_obj value to FI_WAIT_FD.

In the case the wait object to check belongs to a wait set, only the wait set itself needs to be passed into fi_trywait. The fabric resources associated with the wait set do not.

On receiving a return value of -FI_EAGAIN from fi_trywait, an application should read all queued completions and events, and call fi_trywait again before attempting to block. Applications can make use of a fabric poll set to identify completion queues and counters that may require processing.

fi_control

The fi_control call is used to access provider or implementation specific details of a fids that support blocking calls, such as wait sets, completion queues, counters, and event queues. Access to the wait set or fid should be serialized across all calls when fi_control is invoked, as it may redirect the implementation of wait set operations. The following control commands are usable with a wait set or fid.

FI_GETWAIT (void **)

This command allows the user to retrieve the low-level wait object associated with a wait set or fid. The format of the wait set is specified during wait set creation, through the wait set attributes. The fi_control arg parameter should be an address where a pointer to the returned wait object will be written. This should be an 'int *' for FI_WAIT_FD, 'struct fi_mutex_cond' for FI_WAIT_MUTEX_COND, or 'struct fi_wait_pollfd' for FI_WAIT_POLLFD. Support for FI_GETWAIT is provider specific.

FI_GETWAITOBJ (enum fi_wait_obj *)

This command returns the type of wait object associated with a wait set or fid.