On 16/11/2018 16:25, David Gibson wrote:
> On Tue, Nov 13, 2018 at 07:28:12PM +1100, Alexey Kardashevskiy wrote:
>> We already changed NPU API for GPUs to not to call OPAL and the remaining
>> bit is initializing NPU structures.
>>
>> This uses a new QEMU capability which marks NPU-enabled vPHBs as
>> "IBM,npu-vphb" and initializes an NPU structure per vPHB.
>>
>> Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxxxx>
>> ---
>> arch/powerpc/platforms/pseries/pci.c | 6 ++++++
>> 1 file changed, 6 insertions(+)
>>
>> diff --git a/arch/powerpc/platforms/pseries/pci.c b/arch/powerpc/platforms/pseries/pci.c
>> index 41d8a4d..a50d5e4 100644
>> --- a/arch/powerpc/platforms/pseries/pci.c
>> +++ b/arch/powerpc/platforms/pseries/pci.c
>> @@ -29,6 +29,7 @@
>> #include <asm/pci-bridge.h>
>> #include <asm/prom.h>
>> #include <asm/ppc-pci.h>
>> +#include <asm/pci.h>
>> #include "pseries.h"
>>
>> #if 0
>> @@ -237,6 +238,8 @@ static void __init pSeries_request_regions(void)
>>
>> void __init pSeries_final_fixup(void)
>> {
>> + struct pci_controller *hose;
>> +
>> pSeries_request_regions();
>>
>> eeh_probe_devices();
>> @@ -246,6 +249,9 @@ void __init pSeries_final_fixup(void)
>> ppc_md.pcibios_sriov_enable = pseries_pcibios_sriov_enable;
>> ppc_md.pcibios_sriov_disable = pseries_pcibios_sriov_disable;
>> #endif
>> + list_for_each_entry(hose, &hose_list, list_node)
>> + if (of_device_is_compatible(hose->dn, "IBM,npu-vphb"))
>> + pnv_npu2_init(hose);
>
> I take it from this the NPUs are showing up with a compatible property
> that lists the normal PHB value as well as IBM,npu-vphb. Since AIUI
> the NPUs act quite differently from other (real) PHBs this seems
> bogus. Shouldn't they be probed separately?
First, bad naming, will think of better one.
"IBM,npu-vphb" is an extra compatible type for otherwise usual pseries
PHB. The differences are:
1. Initialize an "NPU" (not a NVLink2 bridge but a proper NPU) per a PHB
for context manipulation for a GPU;
2. Kill the default DMA window.
When a GPU is passed to a guest, it looks like:
aik@u1804kvm:~$ lspci
00:00.0 USB controller: NEC Corporation uPD720200 USB 3.0 Host
Controller (rev 03)
00:01.0 Ethernet controller: Red Hat, Inc Virtio network device
00:02.0 3D controller: NVIDIA Corporation GV100 [Tesla V100 SXM2] (rev a1)
00:03.0 Bridge: IBM Device 04ea (rev 01)
00:04.0 Bridge: IBM Device 04ea (rev 01)
So there are:
- one "struct npu" associated with the pseries PHB (not presented in
lspci but there /proc/device-tree/npuphb0/ with link@0 and link@1)
- 2 NVLink2 bridges, presented in lspci.
--
Alexey
