On Fri, Apr 21, 2023 at 03:48:50PM +0800, Binbin Wu wrote: > >On 4/19/2023 11:08 AM, Binbin Wu wrote: >> >> On 4/19/2023 10:30 AM, Chao Gao wrote: >> > On Tue, Apr 04, 2023 at 09:09:21PM +0800, Binbin Wu wrote: >> > > Introduce a new interface untag_addr() to kvm_x86_ops to untag >> > > the metadata >> > >from linear address. Implement LAM version in VMX and dummy version >> > in SVM. >> > > When enabled feature like Intel Linear Address Masking or AMD Upper >> > > Address Ignore, linear address may be tagged with metadata. Linear >> > > address should be checked for modified canonicality and untagged in >> > > instrution emulations or vmexit handlings if LAM or UAI is applicable. >> > > >> > > Introduce untag_addr() to kvm_x86_ops to hide the code related to >> > > vendor >> > > specific details. >> > > - For VMX, LAM version is implemented. >> > > LAM has a modified canonical check when applicable: >> > > * LAM_S48 : [ 1 ][ metadata ][ 1 ] >> > > 63 47 >> > > * LAM_U48 : [ 0 ][ metadata ][ 0 ] >> > > 63 47 >> > > * LAM_S57 : [ 1 ][ metadata ][ 1 ] >> > > 63 56 >> > > * LAM_U57 + 5-lvl paging : [ 0 ][ metadata ][ 0 ] >> > > 63 56 >> > > * LAM_U57 + 4-lvl paging : [ 0 ][ metadata ][ 0...0 ] >> > > 63 56..47 >> > > If LAM is applicable to certain address, untag the metadata bits and >> > > replace them with the value of bit 47 (LAM48) or bit 56 >> > > (LAM57). Later >> > > the untagged address will do legacy canonical check. So that >> > > LAM canonical >> > > check and mask can be covered by "untag + legacy canonical check". >> > > >> > > For cases LAM is not applicable, 'flags' is passed to the interface >> > > to skip untag. >> > The "flags" can be dropped. Callers can simply skip the call of >> > .untag_addr(). >> >> OK. >> >> The "flags" was added for proof of future if such kind of untag is also >> adopted in svm for AMD. >> >> The cases to skip untag are different on the two vendor platforms. >> >> But still, it is able to get the information in __linearize(), I will >> drop the parameter. > >Have a second thought, the flags is still needed to pass to vmx/svm. > >If both implementions set the skip untag flag (SKIP_UNTAG_VMX | >SKIP_UNTAG_SVM) >or neither sets the skip untag flag, __linearize() can decide to call >.untag_addr() or not. > >However, in some case, if only one of the implementation need to set the skip >untag for itself, >in __linearize(), there is no enough information to tell whether to skip the >untag or not. OK. I have no strong preference. An alternative is: We have only one flag. If AMD and Intel differ in some cases, set the flag according to the CPU vendor of vCPUs. if LAM applies to case A while AMD UAI doesn't, do if guest CPU is Intel set SKIP_UNTAG for case B to which UAI applies while LAM doens't, do if guest CPU is AMD set SKIP_UNTAG
