Restore the segment registers and segment base MSRs for amd64 syscall

return path only when neither thread was context switched while
executing syscall code nor syscall explicitely modified LDT or MSRs.

Save segment registers in trap handlers before interrupts are enabled,
to not allow context switches to happen before registers are saved.
Use separated byte in pcb for indication of fast/full return, since
pcb_flags are not synchronized with context switches.

The change puts back syscall microbenchmark numbers that were slowed
down after commit of the support for LDT on amd64.

Reviewed by:	jeff
Tested (and tested, and tested ...) by:	pho
Approved by:	re (kensmith)

3 files changed, 15 insertions, 5 deletions

diff --git a/sys/amd64/ia32/ia32_exception.S b/sys/amd64/ia32/ia32_exception.S
index 76c5d5a..341f00e 100644
--- a/sys/amd64/ia32/ia32_exception.S
+++ b/sys/amd64/ia32/ia32_exception.S
@@ -42,10 +42,16 @@
 	SUPERALIGN_TEXT
 IDTVEC(int0x80_syscall)
 	swapgs
-	sti
 	pushq	$2			/* sizeof "int 0x80" */
 	subq	$TF_ERR,%rsp		/* skip over tf_trapno */
 	movq	%rdi,TF_RDI(%rsp)
+	movq	PCPU(CURPCB),%rdi
+	movb	$0,PCB_FULL_IRET(%rdi)
+	movw	%fs,TF_FS(%rsp)
+	movw	%gs,TF_GS(%rsp)
+	movw	%es,TF_ES(%rsp)
+	movw	%ds,TF_DS(%rsp)
+	sti
 	movq	%rsi,TF_RSI(%rsp)
 	movq	%rdx,TF_RDX(%rsp)
 	movq	%rcx,TF_RCX(%rsp)
@@ -60,10 +66,6 @@ IDTVEC(int0x80_syscall)
 	movq	%r13,TF_R13(%rsp)
 	movq	%r14,TF_R14(%rsp)
 	movq	%r15,TF_R15(%rsp)
-	movw	%fs,TF_FS(%rsp)
-	movw	%gs,TF_GS(%rsp)
-	movw	%es,TF_ES(%rsp)
-	movw	%ds,TF_DS(%rsp)
 	movl	$TF_HASSEGS,TF_FLAGS(%rsp)
 	FAKE_MCOUNT(TF_RIP(%rsp))
 	movq	%rsp, %rdi
diff --git a/sys/amd64/ia32/ia32_reg.c b/sys/amd64/ia32/ia32_reg.c
index 49dd4e2..83f6783 100644
--- a/sys/amd64/ia32/ia32_reg.c
+++ b/sys/amd64/ia32/ia32_reg.c
@@ -125,6 +125,7 @@ set_regs32(struct thread *td, struct reg32 *regs)
 	tp->tf_fs = regs->r_fs;
 	tp->tf_es = regs->r_es;
 	tp->tf_ds = regs->r_ds;
+	td->td_pcb->pcb_full_iret = 1;
 	tp->tf_flags = TF_HASSEGS;
 	tp->tf_rdi = regs->r_edi;
 	tp->tf_rsi = regs->r_esi;
diff --git a/sys/amd64/ia32/ia32_signal.c b/sys/amd64/ia32/ia32_signal.c
index 37e8013..d7c1dd5 100644
--- a/sys/amd64/ia32/ia32_signal.c
+++ b/sys/amd64/ia32/ia32_signal.c
@@ -159,6 +159,7 @@ ia32_get_mcontext(struct thread *td, struct ia32_mcontext *mcp, int flags)
 	ia32_get_fpcontext(td, mcp);
 	mcp->mc_fsbase = td->td_pcb->pcb_fsbase;
 	mcp->mc_gsbase = td->td_pcb->pcb_gsbase;
+	td->td_pcb->pcb_full_iret = 1;
 	return (0);
 }
 
@@ -201,6 +202,7 @@ ia32_set_mcontext(struct thread *td, const struct ia32_mcontext *mcp)
 	tp->tf_rsp = mcp->mc_esp;
 	tp->tf_ss = mcp->mc_ss;
 	td->td_pcb->pcb_flags |= PCB_FULLCTX;
+	td->td_pcb->pcb_full_iret = 1;
 	return (0);
 }
 
@@ -394,6 +396,7 @@ freebsd4_ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	regs->tf_ss = _udatasel;
 	regs->tf_ds = _udatasel;
 	regs->tf_es = _udatasel;
+	td->td_pcb->pcb_full_iret = 1;
 	/* leave user %fs and %gs untouched */
 	PROC_LOCK(p);
 	mtx_lock(&psp->ps_mtx);
@@ -514,6 +517,7 @@ ia32_sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	regs->tf_ss = _udatasel;
 	regs->tf_ds = _udatasel;
 	regs->tf_es = _udatasel;
+	td->td_pcb->pcb_full_iret = 1;
 	/* XXXKIB leave user %fs and %gs untouched */
 	PROC_LOCK(p);
 	mtx_lock(&psp->ps_mtx);
@@ -611,6 +615,7 @@ freebsd4_freebsd32_sigreturn(td, uap)
 	SIG_CANTMASK(td->td_sigmask);
 	signotify(td);
 	PROC_UNLOCK(p);
+	td->td_pcb->pcb_full_iret = 1;
 	return (EJUSTRETURN);
 }
 #endif	/* COMPAT_FREEBSD4 */
@@ -702,6 +707,7 @@ freebsd32_sigreturn(td, uap)
 	SIG_CANTMASK(td->td_sigmask);
 	signotify(td);
 	PROC_UNLOCK(p);
+	td->td_pcb->pcb_full_iret = 1;
 	return (EJUSTRETURN);
 }
 
@@ -747,5 +753,6 @@ ia32_setregs(td, entry, stack, ps_strings)
 	/* Return via doreti so that we can change to a different %cs */
 	pcb->pcb_flags |= PCB_FULLCTX | PCB_32BIT;
 	pcb->pcb_flags &= ~PCB_GS32BIT;
+	td->td_pcb->pcb_full_iret = 1;
 	td->td_retval[1] = 0;
 }