Collect the nastiness for preserving the kernel MSR_GSBASE around the

load_gs() calls into a single place that is less likely to go wrong.

Eliminate the per-process context switching of MSR_GSBASE, because it
should be constant for a single cpu.  Instead, save/restore it during
the loading of the new %gs selector for the new process.

Approved by:	re (amd64/* blanket)

7 files changed, 42 insertions, 36 deletions

diff --git a/sys/amd64/amd64/cpu_switch.S b/sys/amd64/amd64/cpu_switch.S
index 3ca4ecd..9465396 100644
--- a/sys/amd64/amd64/cpu_switch.S
+++ b/sys/amd64/amd64/cpu_switch.S
@@ -106,12 +106,6 @@ ENTRY(cpu_switch)
 	pushfq					/* PSL */
 	popq	PCB_RFLAGS(%r8)
 
-	/* Save kernel %gs.base */
-	movl	$MSR_GSBASE,%ecx
-	rdmsr
-	movl	%eax,PCB_KGSBASE(%r8)
-	movl	%edx,PCB_KGSBASE+4(%r8)
-
 	/* Save userland %fs */
 	movl	$MSR_FSBASE,%ecx
 	rdmsr
@@ -176,12 +170,11 @@ sw1:
 	movl	PCB_DS(%r8),%ds
 	movl	PCB_ES(%r8),%es
 	movl	PCB_FS(%r8),%fs
-	movl	PCB_GS(%r8),%gs
 
-	/* Restore kernel %gs.base */
+	/* Restore userland %gs while preserving kernel gsbase */
 	movl	$MSR_GSBASE,%ecx
-	movl	PCB_KGSBASE(%r8),%eax
-	movl	PCB_KGSBASE+4(%r8),%edx
+	rdmsr
+	movl	PCB_GS(%r8),%gs
 	wrmsr
 
 	/* Restore userland %fs */
diff --git a/sys/amd64/amd64/genassym.c b/sys/amd64/amd64/genassym.c
index b55ef08..3ccad23 100644
--- a/sys/amd64/amd64/genassym.c
+++ b/sys/amd64/amd64/genassym.c
@@ -125,7 +125,6 @@ ASSYM(PCB_RIP, offsetof(struct pcb, pcb_rip));
 ASSYM(PCB_RFLAGS, offsetof(struct pcb, pcb_rflags));
 ASSYM(PCB_FSBASE, offsetof(struct pcb, pcb_fsbase));
 ASSYM(PCB_GSBASE, offsetof(struct pcb, pcb_gsbase));
-ASSYM(PCB_KGSBASE, offsetof(struct pcb, pcb_kgsbase));
 ASSYM(PCB_DS, offsetof(struct pcb, pcb_ds));
 ASSYM(PCB_ES, offsetof(struct pcb, pcb_es));
 ASSYM(PCB_FS, offsetof(struct pcb, pcb_fs));
diff --git a/sys/amd64/amd64/machdep.c b/sys/amd64/amd64/machdep.c
index b0b6ce2..a5d85f1 100644
--- a/sys/amd64/amd64/machdep.c
+++ b/sys/amd64/amd64/machdep.c
@@ -476,21 +476,15 @@ exec_setregs(td, entry, stack, ps_strings)
 {
 	struct trapframe *regs = td->td_frame;
 	struct pcb *pcb = td->td_pcb;
-	u_int64_t pc;
 	
 	wrmsr(MSR_FSBASE, 0);
 	wrmsr(MSR_KGSBASE, 0);	/* User value while we're in the kernel */
 	pcb->pcb_fsbase = 0;
 	pcb->pcb_gsbase = 0;
-	pcb->pcb_kgsbase = rdmsr(MSR_GSBASE);
 	load_ds(_udatasel);
 	load_es(_udatasel);
 	load_fs(_udatasel);
-	critical_enter();
-	pc = rdmsr(MSR_GSBASE);
-	load_gs(_udatasel);	/* Clobbers kernel %GS.base */
-	wrmsr(MSR_GSBASE, pc);
-	critical_exit();
+	load_gs(_udatasel);
 	pcb->pcb_ds = _udatasel;
 	pcb->pcb_es = _udatasel;
 	pcb->pcb_fs = _udatasel;
@@ -1317,7 +1311,6 @@ hammer_time(void)
 	/* setup proc 0's pcb */
 	thread0.td_pcb->pcb_flags = 0; /* XXXKSE */
 	thread0.td_pcb->pcb_cr3 = IdlePML4;
-	thread0.td_pcb->pcb_kgsbase = (u_int64_t)pc;
 	thread0.td_frame = &proc0_tf;
 }
 
diff --git a/sys/amd64/ia32/ia32_sysvec.c b/sys/amd64/ia32/ia32_sysvec.c
index 684677a..5ced18c 100644
--- a/sys/amd64/ia32/ia32_sysvec.c
+++ b/sys/amd64/ia32/ia32_sysvec.c
@@ -241,22 +241,15 @@ ia32_setregs(td, entry, stack, ps_strings)
 {
 	struct trapframe *regs = td->td_frame;
 	struct pcb *pcb = td->td_pcb;
-	u_int64_t pc;
-	register_t s;
 	
 	wrmsr(MSR_FSBASE, 0);
 	wrmsr(MSR_KGSBASE, 0);	/* User value while we're in the kernel */
 	pcb->pcb_fsbase = 0;
 	pcb->pcb_gsbase = 0;
-	pcb->pcb_kgsbase = rdmsr(MSR_GSBASE);
 	load_ds(_udatasel);
 	load_es(_udatasel);
 	load_fs(_udatasel);
-	s = intr_disable();
-	pc = rdmsr(MSR_GSBASE);
-	load_gs(_udatasel);	/* Clobbers kernel %GS.base */
-	wrmsr(MSR_GSBASE, pc);
-	intr_restore(s);
+	load_gs(_udatasel);
 	pcb->pcb_ds = _udatasel;
 	pcb->pcb_es = _udatasel;
 	pcb->pcb_fs = _udatasel;
diff --git a/sys/amd64/include/cpufunc.h b/sys/amd64/include/cpufunc.h
index 21257dd..5d4c0dc 100644
--- a/sys/amd64/include/cpufunc.h
+++ b/sys/amd64/include/cpufunc.h
@@ -475,6 +475,41 @@ load_es(u_int sel)
 	__asm __volatile("movl %0,%%es" : : "rm" (sel));
 }
 
+#ifdef _KERNEL
+/* This is defined in <machine/specialreg.h> but is too painful to get to */
+#ifndef	MSR_FSBASE
+#define	MSR_FSBASE	0xc0000100
+#endif
+static __inline void
+load_fs(u_int sel)
+{
+	register u_int32_t fsbase __asm("ecx");
+
+	/* Preserve the fsbase value across the selector load */
+	fsbase = MSR_FSBASE;
+        __asm __volatile("rdmsr; movl %0,%%fs; wrmsr"
+            : : "rm" (sel), "c" (fsbase) : "eax", "edx");
+}
+
+#ifndef	MSR_GSBASE
+#define	MSR_GSBASE	0xc0000101
+#endif
+static __inline void
+load_gs(u_int sel)
+{
+	register u_int32_t gsbase __asm("ecx");
+
+	/*
+	 * Preserve the gsbase value across the selector load.
+	 * Note that we have to disable interrupts because the gsbase
+	 * being trashed happens to be the kernel gsbase at the time.
+	 */
+	gsbase = MSR_GSBASE;
+        __asm __volatile("pushfq; cli; rdmsr; movl %0,%%gs; wrmsr; popfq"
+            : : "rm" (sel), "c" (gsbase) : "eax", "edx");
+}
+#else
+/* Usable by userland */
 static __inline void
 load_fs(u_int sel)
 {
@@ -486,6 +521,7 @@ load_gs(u_int sel)
 {
 	__asm __volatile("movl %0,%%gs" : : "rm" (sel));
 }
+#endif
 
 /* void lidt(struct region_descriptor *addr); */
 static __inline void
diff --git a/sys/amd64/include/pcb.h b/sys/amd64/include/pcb.h
index 5bfb41f..83d308f 100644
--- a/sys/amd64/include/pcb.h
+++ b/sys/amd64/include/pcb.h
@@ -59,7 +59,6 @@ struct pcb {
 	register_t	pcb_rflags;
 	register_t	pcb_fsbase;
 	register_t	pcb_gsbase;
-	register_t	pcb_kgsbase;
 	u_int32_t	pcb_ds;
 	u_int32_t	pcb_es;
 	u_int32_t	pcb_fs;
diff --git a/sys/compat/ia32/ia32_sysvec.c b/sys/compat/ia32/ia32_sysvec.c
index 684677a..5ced18c 100644
--- a/sys/compat/ia32/ia32_sysvec.c
+++ b/sys/compat/ia32/ia32_sysvec.c
@@ -241,22 +241,15 @@ ia32_setregs(td, entry, stack, ps_strings)
 {
 	struct trapframe *regs = td->td_frame;
 	struct pcb *pcb = td->td_pcb;
-	u_int64_t pc;
-	register_t s;
 	
 	wrmsr(MSR_FSBASE, 0);
 	wrmsr(MSR_KGSBASE, 0);	/* User value while we're in the kernel */
 	pcb->pcb_fsbase = 0;
 	pcb->pcb_gsbase = 0;
-	pcb->pcb_kgsbase = rdmsr(MSR_GSBASE);
 	load_ds(_udatasel);
 	load_es(_udatasel);
 	load_fs(_udatasel);
-	s = intr_disable();
-	pc = rdmsr(MSR_GSBASE);
-	load_gs(_udatasel);	/* Clobbers kernel %GS.base */
-	wrmsr(MSR_GSBASE, pc);
-	intr_restore(s);
+	load_gs(_udatasel);
 	pcb->pcb_ds = _udatasel;
 	pcb->pcb_es = _udatasel;
 	pcb->pcb_fs = _udatasel;