Hi Hans,
On Mon, May 08, 2017 at 03:08:26PM +0200, Hans Verkuil wrote:
> On 05/08/2017 02:56 PM, Sakari Ailus wrote:
...
> >> The USERPTR mode is more dubious. Has this been tested? Can the DMA handle partially
> >> filled pages? (I.e. there must be no requirements such as that the DMA has to start
> >> at a page boundary, since that's not the case with USERPTR).
> >
> > I rememeber this has been discussed before. :-)
> >
> > Most hardware has some limitations on the granularity of the buffer start
> > address, and the drivers still support USERPTR memory. In practice the C
> > library allocated memory is always page aligned if the size is large enough,
> > which is in practice the case for video buffers.
>
> That was not true the last time I checked. I can't remember what the exact
> alignment was, although I do remember that it was different for 32 and 64 bit.
Hmm. This just shows how little I really know about user space. :-I
I just tested this to see how it really works out, and there doesn't seem to
be much alignment at all. I believe my earlier recollection was likely
related to a non-GNU C library (huh!).
>
> I am also pretty sure it was less than 64 bytes. It's been 2 years ago since
> I last looked at this, though.
As in here:
00:26:02 lanttu sailus [~]cat /tmp/foo.c
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
void main() {
unsigned int i;
for (i=0; i < 16; i++)
printf("%x\n",malloc(1024*768*i));
}
00:26:03 lanttu sailus [~]gcc -o /tmp/foo /tmp/foo.c
00:26:06 lanttu sailus [~]/tmp/foo
8604008
f7525008
f73a4008
f7163008
f6e62008
f6aa1008
f6620008
f60df008
f5ade008
f541d008
f4c9c008
f445b008
f3b5a008
f3199008
f2718008
f1bd7008
So posix_memalign() (or memalign()) is needed to allocate page aligned
memory. At least on GNU libc.
--
Regards,
Sakari Ailus
e-mail: sakari.ailus@xxxxxx XMPP: sailus@xxxxxxxxxxxxxx
