SWDEV-452795 - Disable RAS plugin, fix XGMI
RAS plugin loaded rocm-smi which is in conflict with amd-smi library
Main source of grief was the map 'devInfoTypesStrings' that is defined
in both rocm-smi and amd-smi
We assume that rocm-smi would get lazy-loaded by RAS library and
overwrite symbols defined in amd-smi. devInfoTypesStrings in rocm-smi
contains different number of elements, the enums are also different.
RDC relies on amd-smi's enums.
One such enum is kDevGpuMetrics:
rocm-smi: kDevGpuMetrics = 68
amd-smi: kDevGpuMetrics = 75
Example of overlapping map definitions:
$ objdump --dynamic-syms /opt/rocm/lib/libamd_smi.so | grep devInfoTypesStrings
00000000003c4980 g DO .data.rel.ro0000000000000008 Base devInfoTypesStrings
00000000003db830 g DO .bss0000000000000030 Base _ZN3amd3smi6Device19devInfoTypesStringsE
$ objdump --dynamic-syms /opt/rocm/lib/librocm_smi64.so | grep devInfoTypesStrings
00000000003dc590 g DO .bss0000000000000030 Base _ZN3amd3smi6Device19devInfoTypesStringsE
00000000003c9c68 g DO .data.rel.ro0000000000000008 Base devInfoTypesStrings
Change-Id: Ib2f2db32b6abd7ebe84e7807c25581461eb86bae
Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>
[ROCm/rdc commit: d85657e5f2]
This commit is contained in:
@@ -29,7 +29,6 @@ THE SOFTWARE.
|
||||
|
||||
#include "rdc_lib/RdcMetricFetcher.h"
|
||||
#include "rdc_lib/RdcTelemetry.h"
|
||||
#include "rdc_lib/impl/RdcRasLib.h"
|
||||
#include "rdc_lib/impl/RdcSmiLib.h"
|
||||
|
||||
namespace amd {
|
||||
|
||||
Reference in New Issue
Block a user