x86/mm: Use INVPCID for __native_flush_tlb_single()

This uses INVPCID to shoot down individual lines of the user mapping
instead of marking the entire user map as invalid. This
could/might/possibly be faster.

This for sure needs tlb_single_page_flush_ceiling to be redetermined;
esp. since INVPCID is _slow_.

A detailed performance analysis is available here:

  https://lkml.kernel.org/r/3062e486-3539-8a1f-5724-16199420be71@intel.com

[ Peterz: Split out from big combo patch ]
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: Eduardo Valentin <eduval@amazon.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: aliguori@amazon.com
Cc: daniel.gruss@iaik.tugraz.at
Cc: hughd@google.com
Cc: keescook@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/include/asm/cpufeatures.h

@@ -197,6 +197,7 @@
 #define X86_FEATURE_CAT_L3		( 7*32+ 4) /* Cache Allocation Technology L3 */
 #define X86_FEATURE_CAT_L2		( 7*32+ 5) /* Cache Allocation Technology L2 */
 #define X86_FEATURE_CDP_L3		( 7*32+ 6) /* Code and Data Prioritization L3 */
+#define X86_FEATURE_INVPCID_SINGLE	( 7*32+ 7) /* Effectively INVPCID && CR4.PCIDE=1 */
 
 #define X86_FEATURE_HW_PSTATE		( 7*32+ 8) /* AMD HW-PState */
 #define X86_FEATURE_PROC_FEEDBACK	( 7*32+ 9) /* AMD ProcFeedbackInterface */

arch/x86/include/asm/tlbflush.h

@@ -85,6 +85,18 @@ static inline u16 kern_pcid(u16 asid)
 	return asid + 1;
 }
 
+/*
+ * The user PCID is just the kernel one, plus the "switch bit".
+ */
+static inline u16 user_pcid(u16 asid)
+{
+	u16 ret = kern_pcid(asid);
+#ifdef CONFIG_PAGE_TABLE_ISOLATION
+	ret |= 1 << X86_CR3_PTI_SWITCH_BIT;
+#endif
+	return ret;
+}
+
 struct pgd_t;
 static inline unsigned long build_cr3(pgd_t *pgd, u16 asid)
 {
@@ -335,6 +347,8 @@ static inline void __native_flush_tlb_global(void)
 		/*
 		 * Using INVPCID is considerably faster than a pair of writes
 		 * to CR4 sandwiched inside an IRQ flag save/restore.
+		 *
+		 * Note, this works with CR4.PCIDE=0 or 1.
 		 */
 		invpcid_flush_all();
 		return;
@@ -368,7 +382,14 @@ static inline void __native_flush_tlb_single(unsigned long addr)
 	if (!static_cpu_has(X86_FEATURE_PTI))
 		return;
 
+	/*
+	 * Some platforms #GP if we call invpcid(type=1/2) before CR4.PCIDE=1.
+	 * Just use invalidate_user_asid() in case we are called early.
+	 */
+	if (!this_cpu_has(X86_FEATURE_INVPCID_SINGLE))
 		invalidate_user_asid(loaded_mm_asid);
+	else
+		invpcid_flush_one(user_pcid(loaded_mm_asid), addr);
 }
 
 /*

arch/x86/mm/init.c

@@ -203,34 +203,44 @@ static void __init probe_page_size_mask(void)
 
 static void setup_pcid(void)
 {
-#ifdef CONFIG_X86_64
-	if (boot_cpu_has(X86_FEATURE_PCID)) {
+	if (!IS_ENABLED(CONFIG_X86_64))
+		return;
+
+	if (!boot_cpu_has(X86_FEATURE_PCID))
+		return;
+
 	if (boot_cpu_has(X86_FEATURE_PGE)) {
 		/*
-			 * This can't be cr4_set_bits_and_update_boot() --
-			 * the trampoline code can't handle CR4.PCIDE and
-			 * it wouldn't do any good anyway.  Despite the name,
-			 * cr4_set_bits_and_update_boot() doesn't actually
-			 * cause the bits in question to remain set all the
-			 * way through the secondary boot asm.
+		 * This can't be cr4_set_bits_and_update_boot() -- the
+		 * trampoline code can't handle CR4.PCIDE and it wouldn't
+		 * do any good anyway.  Despite the name,
+		 * cr4_set_bits_and_update_boot() doesn't actually cause
+		 * the bits in question to remain set all the way through
+		 * the secondary boot asm.
 		 *
-			 * Instead, we brute-force it and set CR4.PCIDE
-			 * manually in start_secondary().
+		 * Instead, we brute-force it and set CR4.PCIDE manually in
+		 * start_secondary().
 		 */
 		cr4_set_bits(X86_CR4_PCIDE);
+
+		/*
+		 * INVPCID's single-context modes (2/3) only work if we set
+		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
+		 * on systems that have X86_CR4_PCIDE clear, or that have
+		 * no INVPCID support at all.
+		 */
+		if (boot_cpu_has(X86_FEATURE_INVPCID))
+			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
 	} else {
 		/*
-			 * flush_tlb_all(), as currently implemented, won't
-			 * work if PCID is on but PGE is not.  Since that
-			 * combination doesn't exist on real hardware, there's
-			 * no reason to try to fully support it, but it's
-			 * polite to avoid corrupting data if we're on
-			 * an improperly configured VM.
+		 * flush_tlb_all(), as currently implemented, won't work if
+		 * PCID is on but PGE is not.  Since that combination
+		 * doesn't exist on real hardware, there's no reason to try
+		 * to fully support it, but it's polite to avoid corrupting
+		 * data if we're on an improperly configured VM.
 		 */
 		setup_clear_cpu_cap(X86_FEATURE_PCID);
 	}
-	}
-#endif
 }
 
 #ifdef CONFIG_X86_32