x86/paravirt: Fix incorrect virt spinlock setting on bare metal

The kernel can change spinlock behavior when running as a guest. But this
guest-friendly behavior causes performance problems on bare metal.

The kernel uses a static key to switch between the two modes.

In theory, the static key is enabled by default (run in guest mode) and
should be disabled for bare metal (and in some guests that want native
behavior or paravirt spinlock).

A performance drop is reported when running encode/decode workload and
BenchSEE cache sub-workload.

Bisect points to commit ce0a1b6 ("x86/paravirt: Silence unused
native_pv_lock_init() function warning"). When CONFIG_PARAVIRT_SPINLOCKS is
disabled the virt_spin_lock_key is incorrectly set to true on bare
metal. The qspinlock degenerates to test-and-set spinlock, which decreases
the performance on bare metal.

Set the default value of virt_spin_lock_key to false. If booting in a VM,
enable this key. Later during the VM initialization, if other
high-efficient spinlock is preferred (e.g. paravirt-spinlock), or the user
wants the native qspinlock (via nopvspin boot commandline), the
virt_spin_lock_key is disabled accordingly.

This results in the following decision matrix:

X86_FEATURE_HYPERVISOR         Y    Y       Y     N
CONFIG_PARAVIRT_SPINLOCKS      Y    Y       N     Y/N
PV spinlock                    Y    N       N     Y/N

virt_spin_lock_key             N    Y/N     Y     N

Fixes: ce0a1b6 ("x86/paravirt: Silence unused native_pv_lock_init() function warning")
Reported-by: Prem Nath Dey <[email protected]>
Reported-by: Xiaoping Zhou <[email protected]>
Suggested-by: Dave Hansen <[email protected]>
Suggested-by: Qiuxu Zhuo <[email protected]>
Suggested-by: Nikolay Borisov <[email protected]>
Signed-off-by: Chen Yu <[email protected]>
Signed-off-by: Thomas Gleixner <[email protected]>
Reviewed-by: Nikolay Borisov <[email protected]>
Cc: [email protected]
Link: https://lore.kernel.org/all/[email protected]

Author: yu-chen-surf

arch/x86/include/asm/qspinlock.h

@@ -66,13 +66,15 @@ static inline bool vcpu_is_preempted(long cpu)
 
 #ifdef CONFIG_PARAVIRT
 /*
- * virt_spin_lock_key - enables (by default) the virt_spin_lock() hijack.
+ * virt_spin_lock_key - disables by default the virt_spin_lock() hijack.
  *
- * Native (and PV wanting native due to vCPU pinning) should disable this key.
- * It is done in this backwards fashion to only have a single direction change,
- * which removes ordering between native_pv_spin_init() and HV setup.
+ * Native (and PV wanting native due to vCPU pinning) should keep this key
+ * disabled. Native does not touch the key.
+ *
+ * When in a guest then native_pv_lock_init() enables the key first and
+ * KVM/XEN might conditionally disable it later in the boot process again.
  */
-DECLARE_STATIC_KEY_TRUE(virt_spin_lock_key);
+DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
 
 /*
  * Shortcut for the queued_spin_lock_slowpath() function that allows

arch/x86/kernel/paravirt.c

@@ -51,13 +51,12 @@ DEFINE_ASM_FUNC(pv_native_irq_enable, "sti", .noinstr.text);
 DEFINE_ASM_FUNC(pv_native_read_cr2, "mov %cr2, %rax", .noinstr.text);
 #endif
 
-DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
+DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
 
 void __init native_pv_lock_init(void)
 {
-	if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) &&
-	    !boot_cpu_has(X86_FEATURE_HYPERVISOR))
-		static_branch_disable(&virt_spin_lock_key);
+	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
+		static_branch_enable(&virt_spin_lock_key);
 }
 
 static void native_tlb_remove_table(struct mmu_gather *tlb, void *table)