KVM: x86: Fix incorrect memory constraint for FXSAVE in emulator
PATCH 1/1 KVM: x86: Fix incorrect memory constraint for FXSAVE in emulator
The inline asm used to invoke FXSAVE in em_fxsave() and fxregs_fixup() incorrectly specifies the memory operand as read-write ("+m"). FXSAVE does not read from the destination operand; it only writes the current FPU state to memory.
Using a read-write constraint is incorrect and misleading, as it tells the compiler that the previous contents of the buffer are consumed by the instruction. In both cases, the buffer passed to FXSAVE is uninitialized, and marking it as read-write can therefore create a false dependency on uninitialized memory.
Fix the constraint to write-only ("=m") to accurately describe the instruction’s behavior and avoid implying that the buffer is read.
No functional change intended.
--- arch/x86/kvm/emulate.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/x86/kvm/emulate.c +2 -2
@@ -3717,7 +3717,7 @@ static int em_fxsave(struct x86_emulate_ctxt *ctxt)kvm_fpu_get (); - rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_state)); + rc = asm_safe("fxsave %[fx]", , [fx] "=m"(fx_state));kvm_fpu_put ();
@@ -3741,7 +3741,7 @@ static noinline int fxregs_fixup(struct fxregs_state *fx_state,struct fxregs_state fx_tmp;int rc; - rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_tmp)); + rc = asm_safe("fxsave %[fx]", , [fx] "=m"(fx_tmp));memcpy ((void *)fx_state + used_size, (void *)&fx_tmp + used_size,__fxstate_size (16 ) - used_size);
Signature · 1 line
2.53.0
Paolo Bonzini
maintainerQuoted 5 lines · Uros Bizjak
The inline asm used to invoke FXSAVE in
em_fxsave()andfxregs_fixup()incorrectly specifies the memory operand as read-write ("+m"). FXSAVE does not read from the destination operand; it only writes the current FPU state to memory.
Quoted 8 lines
Using a read-write constraint is incorrect and misleading, as it tells the compiler that the previous contents of the buffer are consumed by the instruction. In both cases, the buffer passed to FXSAVE is uninitialized, and marking it as read-write can therefore create a false dependency on uninitialized memory.
Fix the constraint to write-only ("=m") to accurately describe the instruction’s behavior and avoid implying that the buffer is read.
IIRC FXSAVE/FXRSTOR may (at least on some microarchitectures?) leave reserved fields untouched.
Intel suggests writing zeros first, and then the "+m" constraint would be the right one because "=m" would cause the memset to be dead.
Paolo
Quoted 36 lines
No functional change intended.
Signed-off-by Uros Bizjak <[email protected]>Cc Sean Christopherson <[email protected]>Cc Paolo Bonzini <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc Borislav Petkov <[email protected]>Cc Dave Hansen <[email protected]>Cc "H. Peter Anvin" <[email protected]>--- arch/x86/kvm/emulate.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/x86/kvm/emulate.c +2
-2@@ -3717,7 +3717,7 @@ static int em_fxsave(struct x86_emulate_ctxt *ctxt)kvm_fpu_get (); - rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_state)); + rc = asm_safe("fxsave %[fx]", , [fx] "=m"(fx_state));kvm_fpu_put ();@@ -3741,7 +3741,7 @@ static noinline int fxregs_fixup(struct fxregs_state *fx_state,struct fxregs_state fx_tmp;int rc; - rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_tmp)); + rc = asm_safe("fxsave %[fx]", , [fx] "=m"(fx_tmp));memcpy ((void *)fx_state + used_size, (void *)&fx_tmp + used_size,__fxstate_size (16 ) - used_size);
Quoted 13 lines · Paolo Bonzini
Quoted 11 lines · Uros Bizjak
The inline asm used to invoke FXSAVE in
em_fxsave()andfxregs_fixup()incorrectly specifies the memory operand as read-write ("+m"). FXSAVE does not read from the destination operand; it only writes the current FPU state to memory.Using a read-write constraint is incorrect and misleading, as it tells the compiler that the previous contents of the buffer are consumed by the instruction. In both cases, the buffer passed to FXSAVE is uninitialized, and marking it as read-write can therefore create a false dependency on uninitialized memory.
Quoted 8 lines
Quoted 2 lines
Fix the constraint to write-only ("=m") to accurately describe the instruction’s behavior and avoid implying that the buffer is read.
IIRC FXSAVE/FXRSTOR may (at least on some microarchitectures?) leave reserved fields untouched.
Intel suggests writing zeros first, and then the "+m" constraint would be the right one because "=m" would cause the memset to be dead.
Please note that the struct is not initialized before fxsave, so if "+m" is required, the struct should be initialized.
Uros.
Sean Christopherson
maintainerQuoted 17 lines · Uros Bizjak
Quoted 16 lines · Paolo Bonzini
Quoted 13 lines · Uros Bizjak
The inline asm used to invoke FXSAVE in
em_fxsave()andfxregs_fixup()incorrectly specifies the memory operand as read-write ("+m"). FXSAVE does not read from the destination operand; it only writes the current FPU state to memory.Using a read-write constraint is incorrect and misleading, as it tells the compiler that the previous contents of the buffer are consumed by the instruction. In both cases, the buffer passed to FXSAVE is uninitialized, and marking it as read-write can therefore create a false dependency on uninitialized memory.
Fix the constraint to write-only ("=m") to accurately describe the instruction’s behavior and avoid implying that the buffer is read.
Quoted 8 lines
Quoted 5 lines
IIRC FXSAVE/FXRSTOR may (at least on some microarchitectures?) leave reserved fields untouched.
Intel suggests writing zeros first, and then the "+m" constraint would be the right one because "=m" would cause the memset to be dead.
Please note that the struct is not initialized before fxsave, so if "+m" is required, the struct should be initialized.
Regardless of CPU behavior with respect to reserved fields, I believe "+m" is correct and "=m" is wrong, strictly speaking. The SDM very explicitly says:
Bytes 464:511 are available to software use. The processor does not write to bytes 464:511 of an FXSAVE area.
I.e. the entirety of the struct isn't written by FXSAVE, and so using "=m" is technically wrong because those bytes are "read". In practice, it shouldn't matter because fxstate_size() (correctly) truncates the size to a max of 464 bytes, so that KVM-as-the-virutal-CPU honors the architecture and doesn't write to the software-available fields. I.e. those bytes should never truly be read by software.
Given that emulating FXSAVE/FXRSTOR can't possibly be hot paths, explicitly initializing the on-stack structs seems prudent, e.g.
arch/x86/kvm/emulate.c +2 -2
@@ -3708,7 +3708,7 @@ static inline size_t fxstate_size(struct x86_emulate_ctxt *ctxt)*/ static int em_fxsave (struct x86_emulate_ctxt *ctxt) { -struct fxregs_state fx_state; +struct fxregs_state fx_state = {};int rc; rc =check_fxsr (ctxt);
@@ -3738,7 +3738,7 @@ static int em_fxsave(struct x86_emulate_ctxt *ctxt) static noinline int fxregs_fixup(struct fxregs_state *fx_state,const size_t used_size) { -struct fxregs_state fx_tmp; +struct fxregs_state fx_tmp = {};int rc; rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_tmp));