|Skip Navigation Links|
|Exit Print View|
|man pages section 3: Extended Library Functions, Volume 1 Oracle Solaris 11.1 Information Library|
- synchronize local memory with RDMA write on non-coherent memory
cc [ flag… ] file… -ldat [ library… ] #include <dat/udat.h> DAT_RETURN dat_lmr_sync_rdma_write ( IN DAT_IA_HANDLE ia_handle, IN const DAT_LMR_TRIPLET *local_segments, IN DAT_VLEN num_segments )
A handle for an open instance of the IA.
An array of buffer segments.
The number of segments in the local_segments argument.
The dat_lmr_sync_rdma_write() function makes effects of an incoming RDMA Write operation visible to the Consumer. This operation guarantees consistency by locally invalidating the non-coherent cache whose buffer has been populated by remote peer RDMA write operations.
The dat_lmr_sync_rdma_write() function is needed if and only if the Provider attribute specifies that this operation is needed after an incoming RDMA Write operation. The Consumer must call dat_lmr_sync_rdma_write() before reading data from a memory range in this region that was the target of an incoming RDMA Write operation. The dat_lmr_sync_rdma_write() function must be called after the RDMA Write operation completes, and the memory range that was modified by the RDMA Write must be supplied by the caller in the local_ segments array. After this call returns, the Consumer may safely see the modified contents of the memory range. It is permissible to batch synchronizations of multiple RDMA Write operations in a single call by passing a local_segments array that includes all modified memory ranges. The local_segments entries need not contain the same LMR and need not be in the same Protection Zone.
The Consumer must also use dat_lmr_sync_rdma_write() when performing local writes to a memory range that was or will be the target of incoming RDMA writes. After performing the local write, the Consumer must call dat_lmr_sync_rdma_write() before the RDMA Write is initiated. Conversely, after an RDMA Write completes, the Consumer must call dat_lmr_sync_rdma_write() before performing a local write to the same range.
If the Provider attribute specifies that this operation is needed and the Consumer attempts to read from a memory range in an LMR without properly synchronizing using dat_lmr_sync_rdma_write(), the returned contents are undefined. If the Consumer attempts to write to a memory range without properly synchronizing, the contents of the memory range become undefined.
The operation was successful.
The DAT handle is invalid.
One of the parameters is invalid. For example, the address range for a local segment fell outside the boundaries of the corresponding Local Memory Region or the LMR handle was invalid.
Determining when an RDMA Write completes and determining which memory range was modified is the Consumer's responsibility. One possibility is for the RDMA Write initiator to post a Send DTO message after each RDMA Write that identifies the range in the body of the Send. The Consumer at the target of the RDMA Write can receive the message and know when and how to call dat_lmr_sync_rdma_write().
This call ensures that the Provider receives a coherent view of the buffer contents after a subsequent remote RDMA Write operation. After the call completes, the Consumer can be assured that all platform-specific buffer and cache updates have been performed, and that the LMR range has consistency with the Provider hardware. Any subsequent read by the Consumer can void this consistency. The Provider is not required to detect such access.
The action performed on the cache before the RDMA Write depends on the cache type:
I/O noncoherent cache will be flushed.
CPU noncoherent cache will be invalidated.
See attributes(5) for descriptions of the following attributes: