Hello Wanpeng Unfortunately I do not have a simple test case for this issue. The issue I see is with a large non-published project compiled with the latest version of clang. Isn't information provided by QEMU enough? Is there a way to enable additional debug log in KVM? On Mon, Aug 7, 2017 at 3:00 PM, Wanpeng Li <kernellwp@xxxxxxxxx> wrote: > 2017-08-08 2:45 GMT+08:00 Anatol Pomozov <anatol.pomozov@xxxxxxxxx>: >> Hi David >> >> On Mon, Aug 7, 2017 at 9:59 AM, David Matlack <dmatlack@xxxxxxxxxx> wrote: >>> Likely this instruction is accessing an MMIO region (probably a VGA >>> buffer, given "mov $0xb8f0e,%ecx"). This will cause a trap to KVM >>> and KVM will try to emulate the instruction. The instruction doing the >>> MMIO access is an SSE instruction, >> >> It is correct. The emulation error happens in printf() kernel function >> that formats a string and then prints it VGA area. Compilers (both >> clang and gcc) think that SSE instructions are more efficient in >> dealing with strings and that's why I have SSE instructions here. >> >>> which KVM does not have support to >>> emulate, thus the internal error. If you didn't hand write this >>> assembly I would try compiling your kernel with compilation flags to >>> disable SSE code generation and see if that fixes your issue. >> >> Adding "-mno-sse" and removing some floating point functionality in my >> kernel make the emulation error go away. So it confirms your thesis. >> >> Though I need floating point/SSE in my kernel and I can't disable SSE >> completely. Instead I would love to understand why KVM can't use SSE >> with VGA memory? I tested my kernel binary with bare hardware, KVM, >> VMWare, QEMU and only KVM fails. > > Actually the KVM supports SSE emulation in some degree, please post > your testcase here. > > Regards, > Wanpeng Li
