arm64: mm: Always update TCR_EL1 from __cpu_set_tcr_t0sz()

Because only the idmap sets a non-standard T0SZ, __cpu_set_tcr_t0sz()
can check for platforms that need to do this using
__cpu_uses_extended_idmap() before doing its work.

The idmap is only built with enough levels, (and T0SZ bits) to map
its single page.

To allow hibernate, and then kexec to idmap their single page copy
routines, __cpu_set_tcr_t0sz() needs to consider additional users,
who may need a different number of levels/T0SZ-bits to the idmap.
(i.e. VA_BITS may be enough for the idmap, but not hibernate/kexec)

Always read TCR_EL1, and check whether any work needs doing for
this request. __cpu_uses_extended_idmap() remains as it is used
by KVM, whose idmap is also part of the kernel image.

This mostly affects the cpuidle path, where we now get an extra
system register read .

CC: Lorenzo Pieralisi <[email protected]>
CC: Sudeep Holla <[email protected]>
Signed-off-by: James Morse <[email protected]>
Signed-off-by: Pavel Tatashin <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Will Deacon <[email protected]>

arch/arm64/include/asm/mmu_context.h

@@ -81,16 +81,15 @@ static inline bool __cpu_uses_extended_idmap_level(void)
 }
 
 /*
- * Set TCR.T0SZ to its default value (based on VA_BITS)
+ * Ensure TCR.T0SZ is set to the provided value.
  */
 static inline void __cpu_set_tcr_t0sz(unsigned long t0sz)
 {
-	unsigned long tcr;
+	unsigned long tcr = read_sysreg(tcr_el1);
 
-	if (!__cpu_uses_extended_idmap())
+	if ((tcr & TCR_T0SZ_MASK) >> TCR_T0SZ_OFFSET == t0sz)
 		return;
 
-	tcr = read_sysreg(tcr_el1);
 	tcr &= ~TCR_T0SZ_MASK;
 	tcr |= t0sz << TCR_T0SZ_OFFSET;
 	write_sysreg(tcr, tcr_el1);