On Tue, Feb 27, 2024 at 12:20:03AM +0200, Nikolay Borisov wrote: > > > On 27.02.24 г. 0:10 ч., Pawan Gupta wrote: > > On Mon, Feb 26, 2024 at 09:17:30AM +0200, Nikolay Borisov wrote: > > > > diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h > > > > index 262e655..077083e 100644 > > > > --- a/arch/x86/include/asm/nospec-branch.h > > > > +++ b/arch/x86/include/asm/nospec-branch.h > > > > @@ -315,6 +315,17 @@ > > > > #endif > > > > .endm > > > > +/* > > > > + * Macro to execute VERW instruction that mitigate transient data sampling > > > > + * attacks such as MDS. On affected systems a microcode update overloaded VERW > > > > + * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF. > > > > + * > > > > + * Note: Only the memory operand variant of VERW clears the CPU buffers. > > > > + */ > > > > +.macro CLEAR_CPU_BUFFERS > > > > + ALTERNATIVE "", __stringify(verw _ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF > > > > > > Any particular reason why this uses RIP-relative vs an absolute address > > > mode? > > > > Early versions of the series had the VERW arg pointing to the macro > > itself, that is why relative addressing was used. That got changed in a > > later version with all VERW sites pointing to a single memory location. > > > > > I know in our private exchange you said there is no significance but > > > for example older kernels have a missing relocation support in alternatives. > > > This of course can be worked around by slightly changing the logic of the > > > macro which means different kernels will have slightly different macros. > > > > Do you anticipate a problem with that? If yes, I can send a patch to use > > fixed addressing in upstream as well. > > I experienced crashes on older kernels before realizing that the relocation > wasn't resolved correctly by the alternative framework. Instead i simply > changed the macro to jmp 1f, where the next instruction is the verw ( I did > send a backport for 5.4) and it works. Recently there's been a push to make > as much of the kernel assembly as possible PIC so having a rip-relative > addressing helps. Whether that makes any material difference - I cannot say. Ok, sending the patch. > Here's my backport version for reference: > > https://lore.kernel.org/stable/20240226122237.198921-3-nik.borisov@xxxxxxxx/ Below should also solve the problem with less churn: --- diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h index 2aa52cab1e46..ab19c7f1167b 100644 --- a/arch/x86/include/asm/nospec-branch.h +++ b/arch/x86/include/asm/nospec-branch.h @@ -323,7 +323,7 @@ * Note: Only the memory operand variant of VERW clears the CPU buffers. */ .macro CLEAR_CPU_BUFFERS - ALTERNATIVE "", __stringify(verw _ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF + ALTERNATIVE "", __stringify(verw mds_verw_sel), X86_FEATURE_CLEAR_CPU_BUF .endm #else /* __ASSEMBLY__ */