Memory-mapped IO takes a different approach from direct IO to removing the extra data copy performed by normal ("indirect") IO. With mmap-ed IO the application calls the mmap() system call to memory map sg's reserved buffer. The sg driver maintains one reserved buffer per file descriptor. The default size of the reserved buffer is 32 KB and it can be changed with the ioctl(SG_SET_RESERVED_SIZE). The mmap() system call only needs to be called once prior  to doing mmap-ed IO. For more details on the mmap() see Section 7.6. An application indicates that it wants mmap-ed on a SCSI request by setting the SG_FLAG_MMAP_IO value in 'flags'.
Since there is only reserved buffer per sg file descriptor then only one mmap-ed IO command can be active at one time. In order to perform command queuing with mmap-ed IO, an application will need to open() multiple file descriptors to the same SCSI device. With mmap-ed IO the various status values and the sense buffer (if required) are conveyed back to an application in the same fashion as normal ("indirect") IO.
Mmap-ed has very low per command latency since the reserved buffer mapping only needs to be done once per file descriptor. Also the reserved buffer is set up by the sg driver to aid the efficient construction of the internal scatter gather list used by the lower level (adapter) driver for DMA purposes. This tends to be more efficient than the user memory that direct IO requires the sg driver to process into an internal scatter gather list. So on both these counts, mmap-ed IO has the edge over direct IO.
When a write() or ioctl(SG_IO) attempts mmap-ed IO there is no check performed that a prior mmap() system call has been performed. If no mmap() has been issued then random data is written to the device or data read from the device in inaccessible. Also once mmap() has been called on a file descriptor then all subsequent calls to ioctl(SG_SET_RESERVED_SIZE) will yield EBUSY.