xen/arm64: Implement a fast path for handling SMCCC_ARCH_WORKAROUND_1

The function SMCCC_ARCH_WORKAROUND_1 will be called by the guest for
hardening the branch predictor. So we want the handling to be as fast as
possible.

As the mitigation is applied on every guest exit, we can check for the
call before saving all the context and return very early.

For now, only provide a fast path for HVC64 call. Because the code rely
on 2 registers, x0 and x1 are saved in advance.

Signed-off-by: Julien Grall <julien.grall@arm.com>
Reviewed-by: Volodymyr Babchuk <volodymyr.babchuk@epam.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>

xen/arch/arm/arm64/entry.S

@@ -1,6 +1,7 @@
 #include <asm/asm_defns.h>
 #include <asm/regs.h>
 #include <asm/alternative.h>
+#include <asm/smccc.h>
 #include <public/xen.h>
 
 /*
@@ -90,8 +91,12 @@ lr      .req    x30             /* link register */
         .endm
 /*
  * Save state on entry to hypervisor, restore on exit
+ *
+ * save_x0_x1: Does the macro needs to save x0/x1? Defaults to 1
+ * If 0, we rely on the on x0/x1 to have been saved at the correct
+ * position on the stack before.
  */
-        .macro  entry, hyp, compat
+        .macro  entry, hyp, compat, save_x0_x1=1
         sub     sp, sp, #(UREGS_SPSR_el1 - UREGS_LR) /* CPSR, PC, SP, LR */
         push    x28, x29
         push    x26, x27
@@ -107,7 +112,16 @@ lr      .req    x30             /* link register */
         push    x6, x7
         push    x4, x5
         push    x2, x3
+        /*
+         * The caller may already have saved x0/x1 on the stack at the
+         * correct address and corrupt them with another value. Only
+         * save them if save_x0_x1 == 1.
+         */
+        .if \save_x0_x1 == 1
         push    x0, x1
+        .else
+        sub     sp, sp, #16
+        .endif
 
         .if \hyp == 1        /* Hypervisor mode */
 
@@ -200,7 +214,48 @@ hyp_irq:
         exit    hyp=1
 
 guest_sync:
-        entry   hyp=0, compat=0
+        /*
+         * Save x0, x1 in advance
+         */
+        stp     x0, x1, [sp, #-(UREGS_kernel_sizeof - UREGS_X0)]
+
+        /*
+         * x1 is used because x0 may contain the function identifier.
+         * This avoids to restore x0 from the stack.
+         */
+        mrs     x1, esr_el2
+        lsr     x1, x1, #HSR_EC_SHIFT           /* x1 = ESR_EL2.EC */
+        cmp     x1, #HSR_EC_HVC64
+        b.ne    1f                              /* Not a HVC skip fastpath. */
+
+        mrs     x1, esr_el2
+        and     x1, x1, #0xffff                 /* Check the immediate [0:16] */
+        cbnz    x1, 1f                          /* should be 0 for HVC #0 */
+
+        /*
+         * Fastest path possible for ARM_SMCCC_ARCH_WORKAROUND_1.
+         * The workaround has already been applied on the exception
+         * entry from the guest, so let's quickly get back to the guest.
+         *
+         * Note that eor is used because the function identifier cannot
+         * be encoded as an immediate for cmp.
+         */
+        eor     w0, w0, #ARM_SMCCC_ARCH_WORKAROUND_1_FID
+        cbnz    w0, 1f
+
+        /*
+         * Clobber both x0 and x1 to prevent leakage. Note that thanks
+         * the eor, x0 = 0.
+         */
+        mov     x1, xzr
+        eret
+
+1:
+        /*
+         * x0/x1 may have been scratch by the fast path above, so avoid
+         * to save them.
+         */
+        entry   hyp=0, compat=0, save_x0_x1=0
         /*
          * The vSError will be checked while SKIP_SYNCHRONIZE_SERROR_ENTRY_EXIT
          * is not set. If a vSError took place, the initial exception will be

xen/include/asm-arm/processor.h

@@ -306,6 +306,8 @@
 #define HDCR_TPM        (_AC(1,U)<<6)           /* Trap Performance Monitors accesses */
 #define HDCR_TPMCR      (_AC(1,U)<<5)           /* Trap PMCR accesses */
 
+#define HSR_EC_SHIFT                26
+
 #define HSR_EC_UNKNOWN              0x00
 #define HSR_EC_WFI_WFE              0x01
 #define HSR_EC_CP15_32              0x03