Signed-off-by: Arthur Kepner <firstname.lastname@example.org>
DMA-mapping.txt | 25 ++++++++++++++++++++++++-
1 file changed, 24 insertions(+), 1 deletion(-)
diff --git a/Documentation/DMA-mapping.txt b/Documentation/DMA-mapping.txt
index e07f253..32f88e5 100644
@@ -745,7 +745,30 @@ to "Closing".
2.5., page+offset is always used, and the "address" field has been
-2) More to come...
+2) Platforms that permit DMA reordering
+ At least one platform (SGI SN2) allows DMA to be reordered between
+ device and host memory. This reordering can happen in the NUMA
+ interconnect, and it usually results in correct operation and improved
+ performance. In some situations it may be necessary to explicitly
+ synchronize DMA from the device. For example, if the device does a DMA
+ write to indicate that data is ready, the DMA of the "completion
+ indication" can race with DMA of data. In this situation,
+ pci_dma_flags_set_dmaflush() can be used to associate a "dmaflush"
+ attribute with a memory region. The dmaflush attribute will cause in-
+ flight DMA from the device to be flushed to host memory when the
+ associated memory region is written. (On platforms that don't permit
+ DMA reordering, this is a no-op.)
+ For example, you might map the memory region used for completion
+ indications as follows:
+ int count, flags = pci_dma_flags_set_dmaflush(DMA_BIDIRECTIONAL);
+ count = dma_map_sg(dev, sglist, nents, flags);
+3) More to come...