x86/entry/64/compat: After SYSENTER, move STI after the NT fixup
authorAndy Lutomirski <luto@kernel.org>
Tue, 6 Oct 2015 00:47:55 +0000 (17:47 -0700)
committerIngo Molnar <mingo@kernel.org>
Wed, 7 Oct 2015 09:34:08 +0000 (11:34 +0200)
We eventually want to make it all the way into C code before
enabling interrupts.  We need to rework our flags handling
slightly to delay enabling interrupts.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/r/35d24d2a9305da3182eab7b2cdfd32902e90962c.1444091584.git.luto@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/entry/entry_64_compat.S

index e2cca89c1ed33a9f4b5ee00d37c18b39f9c10600..bc678f0c3c9174a2d3d3599bd47d9a8aca42e453 100644 (file)
@@ -58,14 +58,9 @@ ENDPROC(native_usergs_sysret32)
  * with the int 0x80 path.
  */
 ENTRY(entry_SYSENTER_compat)
-       /*
-        * Interrupts are off on entry.
-        * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
-        * it is too small to ever cause noticeable irq latency.
-        */
+       /* Interrupts are off on entry. */
        SWAPGS_UNSAFE_STACK
        movq    PER_CPU_VAR(cpu_current_top_of_stack), %rsp
-       ENABLE_INTERRUPTS(CLBR_NONE)
 
        /* Zero-extending 32-bit regs, do not remove */
        movl    %ebp, %ebp
@@ -76,7 +71,16 @@ ENTRY(entry_SYSENTER_compat)
        /* Construct struct pt_regs on stack */
        pushq   $__USER32_DS            /* pt_regs->ss */
        pushq   %rbp                    /* pt_regs->sp */
-       pushfq                          /* pt_regs->flags */
+
+       /*
+        * Push flags.  This is nasty.  First, interrupts are currently
+        * off, but we need pt_regs->flags to have IF set.  Second, even
+        * if TF was set when SYSENTER started, it's clear by now.  We fix
+        * that later using TIF_SINGLESTEP.
+        */
+       pushfq                          /* pt_regs->flags (except IF = 0) */
+       orl     $X86_EFLAGS_IF, (%rsp)  /* Fix saved flags */
+
        pushq   $__USER32_CS            /* pt_regs->cs */
        pushq   %r10                    /* pt_regs->ip = thread_info->sysenter_return */
        pushq   %rax                    /* pt_regs->orig_ax */
@@ -92,11 +96,21 @@ ENTRY(entry_SYSENTER_compat)
         * Sysenter doesn't filter flags, so we need to clear NT
         * ourselves.  To save a few cycles, we can check whether
         * NT was set instead of doing an unconditional popfq.
+        * This needs to happen before enabling interrupts so that
+        * we don't get preempted with NT set.
         */
        testl   $X86_EFLAGS_NT, EFLAGS(%rsp)
        jnz     sysenter_fix_flags
 sysenter_flags_fixed:
 
+       /*
+        * Re-enable interrupts.  IRQ tracing already thinks that IRQs are
+        * on (since we treat user mode as having IRQs on), and the
+        * prologue above is too short for it to be worth adding a
+        * tracing round trip.
+        */
+       ENABLE_INTERRUPTS(CLBR_NONE)
+
        /*
         * No need to do an access_ok() check here because RBP has been
         * 32-bit zero extended:
@@ -244,7 +258,7 @@ sysexit_audit:
 #endif
 
 sysenter_fix_flags:
-       pushq   $(X86_EFLAGS_IF|X86_EFLAGS_FIXED)
+       pushq   $X86_EFLAGS_FIXED
        popfq
        jmp     sysenter_flags_fixed