DDI_DMA_SYNC(9F) Kernel Functions for Drivers DDI_DMA_SYNC(9F)


ddi_dma_sync - synchronize CPU and I/O views of memory


#include <sys/ddi.h>
#include <sys/sunddi.h>

int ddi_dma_sync(ddi_dma_handle_t handle, off_t offset,
size_t length, uint_t type);


illumos DDI specific (illumos DDI).


The handle filled in by a call to ddi_dma_alloc_handle(9F).

The offset into the object described by the handle.

The length, in bytes, of the area to synchronize. When length
is zero, the entire range starting from offset to the end of
the object has the requested operation applied to it.

Indicates the caller's desire about what view of the memory
object to synchronize. The possible values are


The ddi_dma_sync() function is used to selectively synchronize either a
DMA device's or a CPU's view of a memory object that has DMA resources
allocated for I/O . This may involve operations such as flushes of CPU or
I/O caches, as well as other more complex operations such as stalling
until hardware write buffers have drained.

This function need only be called under certain circumstances. When
resources are allocated for DMA using ddi_dma_addr_bind_handle() or
ddi_dma_buf_bind_handle(), an implicit ddi_dma_sync() is done. When DMA
resources are deallocated using ddi_dma_unbind_handle(9F), an implicit
ddi_dma_sync() is done. However, at any time between DMA resource
allocation and deallocation, if the memory object has been modified by
either the DMA device or a CPU and you wish to ensure that the change is
noticed by the party that did not do the modifying, a call to
ddi_dma_sync() is required. This is true independent of any attributes of
the memory object including, but not limited to, whether or not the
memory was allocated for consistent mode I/O (see ddi_dma_mem_alloc(9F))
or whether or not DMA resources have been allocated for consistent mode
I/O (see ddi_dma_addr_bind_handle(9F) or ddi_dma_buf_bind_handle(9F)).

If a consistent view of the memory object must be ensured between the
time DMA resources are allocated for the object and the time they are
deallocated, you must call ddi_dma_sync() to ensure that either a CPU or
a DMA device has such a consistent view.

What to set type to depends on the view you are trying to ensure
consistency for. If the memory object is modified by a CPU, and the
object is going to be read by the DMA engine of the device, use
DDI_DMA_SYNC_FORDEV. This ensures that the device's DMA engine sees any
changes that a CPU has made to the memory object. If the DMA engine for
the device has written to the memory object, and you are going to read
(with a CPU) the object (using an extant virtual address mapping that you
have to the memory object), use DDI_DMA_SYNC_FORCPU. This ensures that a
CPU's view of the memory object includes any changes made to the object
by the device's DMA engine. If you are only interested in the kernel's
view (kernel-space part of the CPU's view) you may use
DDI_DMA_SYNC_FORKERNEL. This gives a hint to the system--that is, if it
is more economical to synchronize the kernel's view only, then do so;
otherwise, synchronize for CPU.


The ddi_dma_sync() function returns:

Caches are successfully flushed.

The address range to be flushed is out of the address
range established by ddi_dma_addr_bind_handle(9F) or


The ddi_dma_sync() function can be called from user, interrupt, or kernel


ddi_dma_addr_bind_handle(9F), ddi_dma_alloc_handle(9F),
ddi_dma_buf_bind_handle(9F), ddi_dma_mem_alloc(9F),

Writing Device Drivers

January 16, 2006 DDI_DMA_SYNC(9F)