It's not 'direct':
Ishtar:dev/shm# dd if=/dev/zero of=4G bs=2G count=2
dd: warning: partial read (2147479552 bytes); suggest iflag=fullblock
0+2 records in
0+2 records out
4294959104 bytes (4.3 GB) copied, 4.29234 s, 1.0 GB/s
Ishtar:dev/shm# dd if=/dev/zero of=4G bs=2G count=2 iflag=fullblock
2+0 records in
2+0 records out
4294967296 bytes (4.3 GB) copied, 5.41603 s, 793 MB/s
---
 From /dev/shm (file to file)
Ishtar:dev/shm# dd if=4G of=4Ga bs=2G count=2
dd: warning: partial read (2147479552 bytes); suggest iflag=fullblock
0+2 records in
0+2 records out
4294959104 bytes (4.3 GB) copied, 5.52481 s, 777 MB/s
-----------------


Hey guys, this is still a bug though:

Ishtar:dev/shm# dd if=4G of=4Ga bs=4G count=1
0+1 records in
0+1 records out
2147479552 bytes (2.1 GB) copied, 4.0274 s, 533 MB/s
(note no error message....)




Paul Eggert wrote:
> On 10/12/2012 02:14 AM, Pádraig Brady wrote:
>   
>> if you specify ibs and obs seperately,
>> then you will get writes of the size you requested at least.
>>     
>
> Won't there be similar problems with the write system
> call too?  Perhaps not on Linda's system, but on other
> systems.  Once the buffer size exceeds 2 GB or so,
> things get pretty dicey in the wild, not due to any
> problem in dd itself, but due to the kernels or
> file systems that dd relies on.
>