[PATCH 4.9 30/78] x86/switch_to/64: Rewrite FS/GS switching yet again to fix AMD CPUs

[Greg Kroah-Hartman]

4.9-stable review patch.  If anyone has any objections, please let me know.

------------------

From: Andy Lutomirski <luto@xxxxxxxxxx>

commit e137a4d8f4dd2e277e355495b6b2cb241a8693c3 upstream.

Switching FS and GS is a mess, and the current code is still subtly
wrong: it assumes that "Loading a nonzero value into FS sets the
index and base", which is false on AMD CPUs if the value being
loaded is 1, 2, or 3.

(The current code came from commit 3e2b68d752c9 ("x86/asm,
sched/x86: Rewrite the FS and GS context switch code"), which made
it better but didn't fully fix it.)

Rewrite it to be much simpler and more obviously correct.  This
should fix it fully on AMD CPUs and shouldn't adversely affect
performance.

Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxxxx>
Cc: Borislav Petkov <bpetkov@xxxxxxx>
Cc: Brian Gerst <brgerst@xxxxxxxxx>
Cc: Chang Seok <chang.seok.bae@xxxxxxxxx>
Cc: Denys Vlasenko <dvlasenk@xxxxxxxxxx>
Cc: H. Peter Anvin <hpa@xxxxxxxxx>
Cc: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>

---
 arch/x86/kernel/process_64.c |  227 +++++++++++++++++++++++--------------------
 1 file changed, 122 insertions(+), 105 deletions(-)

--- a/arch/x86/kernel/process_64.c
+++ b/arch/x86/kernel/process_64.c
@@ -136,6 +136,123 @@ void release_thread(struct task_struct *
 	}
 }
 
+enum which_selector {
+	FS,
+	GS
+};
+
+/*
+ * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
+ * not available.  The goal is to be reasonably fast on non-FSGSBASE systems.
+ * It's forcibly inlined because it'll generate better code and this function
+ * is hot.
+ */
+static __always_inline void save_base_legacy(struct task_struct *prev_p,
+					     unsigned short selector,
+					     enum which_selector which)
+{
+	if (likely(selector == 0)) {
+		/*
+		 * On Intel (without X86_BUG_NULL_SEG), the segment base could
+		 * be the pre-existing saved base or it could be zero.  On AMD
+		 * (with X86_BUG_NULL_SEG), the segment base could be almost
+		 * anything.
+		 *
+		 * This branch is very hot (it's hit twice on almost every
+		 * context switch between 64-bit programs), and avoiding
+		 * the RDMSR helps a lot, so we just assume that whatever
+		 * value is already saved is correct.  This matches historical
+		 * Linux behavior, so it won't break existing applications.
+		 *
+		 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
+		 * report that the base is zero, it needs to actually be zero:
+		 * see the corresponding logic in load_seg_legacy.
+		 */
+	} else {
+		/*
+		 * If the selector is 1, 2, or 3, then the base is zero on
+		 * !X86_BUG_NULL_SEG CPUs and could be anything on
+		 * X86_BUG_NULL_SEG CPUs.  In the latter case, Linux
+		 * has never attempted to preserve the base across context
+		 * switches.
+		 *
+		 * If selector > 3, then it refers to a real segment, and
+		 * saving the base isn't necessary.
+		 */
+		if (which == FS)
+			prev_p->thread.fsbase = 0;
+		else
+			prev_p->thread.gsbase = 0;
+	}
+}
+
+static __always_inline void save_fsgs(struct task_struct *task)
+{
+	savesegment(fs, task->thread.fsindex);
+	savesegment(gs, task->thread.gsindex);
+	save_base_legacy(task, task->thread.fsindex, FS);
+	save_base_legacy(task, task->thread.gsindex, GS);
+}
+
+static __always_inline void loadseg(enum which_selector which,
+				    unsigned short sel)
+{
+	if (which == FS)
+		loadsegment(fs, sel);
+	else
+		load_gs_index(sel);
+}
+
+static __always_inline void load_seg_legacy(unsigned short prev_index,
+					    unsigned long prev_base,
+					    unsigned short next_index,
+					    unsigned long next_base,
+					    enum which_selector which)
+{
+	if (likely(next_index <= 3)) {
+		/*
+		 * The next task is using 64-bit TLS, is not using this
+		 * segment at all, or is having fun with arcane CPU features.
+		 */
+		if (next_base == 0) {
+			/*
+			 * Nasty case: on AMD CPUs, we need to forcibly zero
+			 * the base.
+			 */
+			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
+				loadseg(which, __USER_DS);
+				loadseg(which, next_index);
+			} else {
+				/*
+				 * We could try to exhaustively detect cases
+				 * under which we can skip the segment load,
+				 * but there's really only one case that matters
+				 * for performance: if both the previous and
+				 * next states are fully zeroed, we can skip
+				 * the load.
+				 *
+				 * (This assumes that prev_base == 0 has no
+				 * false positives.  This is the case on
+				 * Intel-style CPUs.)
+				 */
+				if (likely(prev_index | next_index | prev_base))
+					loadseg(which, next_index);
+			}
+		} else {
+			if (prev_index != next_index)
+				loadseg(which, next_index);
+			wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
+			       next_base);
+		}
+	} else {
+		/*
+		 * The next task is using a real segment.  Loading the selector
+		 * is sufficient.
+		 */
+		loadseg(which, next_index);
+	}
+}
+
 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
 		unsigned long arg, struct task_struct *p, unsigned long tls)
 {
@@ -273,7 +390,6 @@ __switch_to(struct task_struct *prev_p,
 	struct fpu *next_fpu = &next->fpu;
 	int cpu = smp_processor_id();
 	struct tss_struct *tss = &per_cpu(cpu_tss, cpu);
-	unsigned prev_fsindex, prev_gsindex;
 	fpu_switch_t fpu_switch;
 
 	fpu_switch = switch_fpu_prepare(prev_fpu, next_fpu, cpu);
@@ -283,8 +399,7 @@ __switch_to(struct task_struct *prev_p,
 	 *
 	 * (e.g. xen_load_tls())
 	 */
-	savesegment(fs, prev_fsindex);
-	savesegment(gs, prev_gsindex);
+	save_fsgs(prev_p);
 
 	/*
 	 * Load TLS before restoring any segments so that segment loads
@@ -323,108 +438,10 @@ __switch_to(struct task_struct *prev_p,
 	if (unlikely(next->ds | prev->ds))
 		loadsegment(ds, next->ds);
 
-	/*
-	 * Switch FS and GS.
-	 *
-	 * These are even more complicated than DS and ES: they have
-	 * 64-bit bases are that controlled by arch_prctl.  The bases
-	 * don't necessarily match the selectors, as user code can do
-	 * any number of things to cause them to be inconsistent.
-	 *
-	 * We don't promise to preserve the bases if the selectors are
-	 * nonzero.  We also don't promise to preserve the base if the
-	 * selector is zero and the base doesn't match whatever was
-	 * most recently passed to ARCH_SET_FS/GS.  (If/when the
-	 * FSGSBASE instructions are enabled, we'll need to offer
-	 * stronger guarantees.)
-	 *
-	 * As an invariant,
-	 * (fsbase != 0 && fsindex != 0) || (gsbase != 0 && gsindex != 0) is
-	 * impossible.
-	 */
-	if (next->fsindex) {
-		/* Loading a nonzero value into FS sets the index and base. */
-		loadsegment(fs, next->fsindex);
-	} else {
-		if (next->fsbase) {
-			/* Next index is zero but next base is nonzero. */
-			if (prev_fsindex)
-				loadsegment(fs, 0);
-			wrmsrl(MSR_FS_BASE, next->fsbase);
-		} else {
-			/* Next base and index are both zero. */
-			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
-				/*
-				 * We don't know the previous base and can't
-				 * find out without RDMSR.  Forcibly clear it.
-				 */
-				loadsegment(fs, __USER_DS);
-				loadsegment(fs, 0);
-			} else {
-				/*
-				 * If the previous index is zero and ARCH_SET_FS
-				 * didn't change the base, then the base is
-				 * also zero and we don't need to do anything.
-				 */
-				if (prev->fsbase || prev_fsindex)
-					loadsegment(fs, 0);
-			}
-		}
-	}
-	/*
-	 * Save the old state and preserve the invariant.
-	 * NB: if prev_fsindex == 0, then we can't reliably learn the base
-	 * without RDMSR because Intel user code can zero it without telling
-	 * us and AMD user code can program any 32-bit value without telling
-	 * us.
-	 */
-	if (prev_fsindex)
-		prev->fsbase = 0;
-	prev->fsindex = prev_fsindex;
-
-	if (next->gsindex) {
-		/* Loading a nonzero value into GS sets the index and base. */
-		load_gs_index(next->gsindex);
-	} else {
-		if (next->gsbase) {
-			/* Next index is zero but next base is nonzero. */
-			if (prev_gsindex)
-				load_gs_index(0);
-			wrmsrl(MSR_KERNEL_GS_BASE, next->gsbase);
-		} else {
-			/* Next base and index are both zero. */
-			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
-				/*
-				 * We don't know the previous base and can't
-				 * find out without RDMSR.  Forcibly clear it.
-				 *
-				 * This contains a pointless SWAPGS pair.
-				 * Fixing it would involve an explicit check
-				 * for Xen or a new pvop.
-				 */
-				load_gs_index(__USER_DS);
-				load_gs_index(0);
-			} else {
-				/*
-				 * If the previous index is zero and ARCH_SET_GS
-				 * didn't change the base, then the base is
-				 * also zero and we don't need to do anything.
-				 */
-				if (prev->gsbase || prev_gsindex)
-					load_gs_index(0);
-			}
-		}
-	}
-	/*
-	 * Save the old state and preserve the invariant.
-	 * NB: if prev_gsindex == 0, then we can't reliably learn the base
-	 * without RDMSR because Intel user code can zero it without telling
-	 * us and AMD user code can program any 32-bit value without telling
-	 * us.
-	 */
-	if (prev_gsindex)
-		prev->gsbase = 0;
-	prev->gsindex = prev_gsindex;
+	load_seg_legacy(prev->fsindex, prev->fsbase,
+			next->fsindex, next->fsbase, FS);
+	load_seg_legacy(prev->gsindex, prev->gsbase,
+			next->gsindex, next->gsbase, GS);
 
 	switch_fpu_finish(next_fpu, fpu_switch);