IBGDA PeerMappingOverride -> Bug 1630288

[frdpnlxp]

Hi,
I'm considering setting option NVreg_RegistryDwords="PeerMappingOverride=1;" on nvidia.ko, as suggested by NVSHMEM for IBGDA support (GPU Direct Async). This is currently disabled by default, but accessible for root.

Browsing the source code for information I noticed several worrying lines :)
For example, in src/nvidia/arch/nvalloc/unix/src/osmemdesc.c:

 452     if (FLD_TEST_DRF(OS02, _FLAGS, _PEER_MAP_OVERRIDE, _REQUIRED, flags))
 453     {
 454         //
 455         // Don't allow MMIO mappings for unprivileged users
 456         // This is a temporary WAR for bug 1630288 "[PeerSync] threat related
 457         // to GPU"
 458         //
 459         KernelBif *pKernelBif = GPU_GET_KERNEL_BIF(pGpu);
 460         if (!pKernelBif->peerMappingOverride && !osIsAdministrator())
 461         {
 462             NV_PRINTF(LEVEL_ERROR,
 463                       "%s(): permission denied, allowPeermapping=%d\n",
 464                       __FUNCTION__, pKernelBif->peerMappingOverride);                                                                                       
 465             return NV_ERR_INSUFFICIENT_PERMISSIONS;
 466         }
 467     }

I cannot find information on Bug 1630288 online. But it seems the work around is to restrict the option, which is a little worrying.

My question is what are the details for this bug? Specifically, what are the implications, risk surface, etc.

Thank you for any information!