Hi Ohad,
On Fri, Mar 16, 2012 at 4:57 PM, Michal Simek <monstr@xxxxxxxxx> wrote:
In previous version I had carverout first which means that code was copyied
to physical address 0x0 and then vrigns were allocated. Current code
allocate vring first and then RTOS is not able to run from zero address.
In general, depending on specific allocation ordering is error prone
and you shouldn't do it.
Yes, I agree with you. It was in v1.
For a given device, we treat memory allocated with
dma_alloc_coherent() as homogeneous. We care about its utilization and
alignments of allocations rather than their specific location in the
total region.
For zynq platform is very important to use addresses from 0 for rtos.
The main reason why it is so important is ARM reset vectors.
If that's not the case, and there's special memory you want assigned
with a specific purpose, than it's recommended to let the code make
sure it happens explicitly, and not rely on implicit order of
allocation.
Sure.
Today there's no way to assign several memory regions with a specific
device, but a reasonable workaround that's being used is creating
several sub-devices with the special memory, and then looking them up
with stuff like device_for_each_child() or device_find_child().
About your specific issue: the current allocation scheme was chosen
for simplicity, but I do plan to change it back by postponing the
vring dma allocation until a remote processor is really powered on.
It's not a lot of memory, but this way we'd only allocate it when
really needed (so if it's CMA, others could use it).
IOW: you can probably just wait a bit until this patch is ready and
take it into your tree. It will most likely bring back the behavior you
need :)
OK. I will use v1 where allocation is based on order in resource table.
It is not ideal but it is working.
From my point of view will be the best to call dma_alloc_coherent only once
which will allocate the whole chunk of memory which is required for "RTOS".
My resource table is easy:
#define TEXT_BASE 0
struct resource_table __resource resources = {
1, /* we're the first version that implements this */
3, /* number of entries in the table */
0, 0, /* reserved, must be zero */
/* offsets to entries */
{
offsetof(struct resource_table, text_cout),
offsetof(struct resource_table, rpmsg_vdev),
offsetof(struct resource_table, trace),
},
{
TYPE_CARVEOUT, 0, 0, SZ_1M, 0, 0, "MEM_DATA",
},
{
TYPE_VDEV, VIRTIO_ID_RPMSG, 0,
RPMSG_IPU_C0_FEATURES, 0, 0, 0, 2, { 0, 0 },
/* no config data */
},
/* the two vrings */
{ TEXT_BASE + SZ_1M, 4096, 256, 1, 0 },
{ TEXT_BASE + SZ_1M + 0x4000, 4096, 256, 2, 0 },
{
TYPE_TRACE, TEXT_BASE + 0x40000, 0x8000, 0, "trace",
},
};
But as you see I need to keep that offset structure which is
that text/data starts at address 0, vrings0 on 0x100000, vrigns1 on 0x104000
and trace buffer 0x140000.
If dma_alloc_coherent is called just once with size taken from resource table.
(For my case for example 2MB=0x200000) and offset are driven by remoteproc then
it should just work.
Thanks,
Michal
--
Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
Maintainer of Linux kernel 2.6 Microblaze Linux - http://www.monstr.eu/fdt/
Microblaze U-BOOT custodian
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
