Although, I made sure to not break memory isolatio with the
following
workaround:
When booting a root cell for Jailhouse, you would typically carveout
memory for the
inmate cell. I have defined the cell configs such that, in the root
cell
config, RAM region for inmate is
NOT marked with MEM_DMA, this way it never gets mapped in PVU.
When creating cell, root cell maps the inmate RAM loadable region,
here
that memory is not
mapped in IO space.
---> Limitation of this is that you cannot DMA copy the images in
the
loadable sections,
which we are not doing anyways
When destroying the cell, Jailhouse should map the memory back to
the
root cell.
Here, again, the inmate RAM region gets ignored in IO mapping
because of
lacking flag MEM_DMA
cell_create and cell_destroy work in regression, tested
successfully.
Please add at least a test to the unmap function that warns when a
config is violating the hardware constraints. It's still not clear to
me, though, how well that goes with changing inmate cell configs.
Let me explain via the code
root cell config for j721e-evm:
1 /* RAM - first bank*/ {
.phys_start = 0x80000000,
.virt_start = 0x80000000,
.size = 0x80000000,
.flags = JAILHOUSE_MEM_READ | JAILHOUSE_MEM_WRITE |
JAILHOUSE_MEM_EXECUTE | JAILHOUSE_MEM_DMA |
JAILHOUSE_MEM_LOADABLE,
},
2 /* RAM - second bank */ {
.phys_start = 0x880000000,
.virt_start = 0x880000000,
.size = 0x1fa00000,
.flags = JAILHOUSE_MEM_READ | JAILHOUSE_MEM_WRITE |
JAILHOUSE_MEM_EXECUTE | JAILHOUSE_MEM_DMA |
JAILHOUSE_MEM_LOADABLE,
},
3 /* RAM - reserved for ivshmem and baremetal apps */ {
.phys_start = 0x89fe00000,
.virt_start = 0x89fe00000,
.size = 0x200000,
.flags = JAILHOUSE_MEM_READ | JAILHOUSE_MEM_WRITE |
JAILHOUSE_MEM_EXECUTE | JAILHOUSE_MEM_LOADABLE,
},
4 /* RAM - reserved for inmate */ {
.phys_start = 0x8a0000000,
.virt_start = 0x8a0000000,
.size = 0x60000000,
.flags = JAILHOUSE_MEM_READ | JAILHOUSE_MEM_WRITE |
JAILHOUSE_MEM_EXECUTE | JAILHOUSE_MEM_LOADABLE,
},
Here, note that all of 1,2,34 gets mapped in CPU MMU, but only 1,2
gets
mapped in PVU
inmate cell config for j721e-evm-linux-demo:
5 /* RAM. */ {
.phys_start = 0x8a0000000,
.virt_start = 0x8a0000000,
.size = 0x60000000,
.flags = JAILHOUSE_MEM_READ | JAILHOUSE_MEM_WRITE |
JAILHOUSE_MEM_EXECUTE | JAILHOUSE_MEM_DMA |
JAILHOUSE_MEM_LOADABLE,
},
* When enabling jailhouse:
In ideal world, all of the 1,2,3,4(same as 5) should be mapped in
CPU MMU and PVU
With current config, only 1,2,3 is mapped. root cell kernel
continues without any trouble
Since the inmate memory is reserved, no driver attempts
DMA to
that region, no faults seen
* When creating inmate cell:
In ideal world, the IO mapping from PVU should be removed from
root
cell stream ID and to be added in the inmate cell stream ID
With current config, unmap return 0 because nothing was ever
mapped
* When loading images (SET_LOADABLE):
In ideal world, loadable regions should be mapped in the PVU map
for root cell streamID
Since the MEM_DMA is missing, PVU unit skips this chunk and never
maps to root cell
If you DMA copy the images to the root cell view of inmate
loadable
memory, there will be faults
We do not do this currently (I believe we CPU copy the images)
Correct me if I am wrong here
* When starting cell
Again, ideally the mapping should be removed from root cell and
added to inmate cell
unmap returns 0 becasuse it was never mapped
pvu_iommu_program_entries called in inmate 2nd time does
nothing if
the pvu_tlb_is_enabled returns true
Nowhere, PVU unit reprograms the memory map to add or remove entires.
Because it doesn't have to do.
