kernel/x86: Fix double fault stack allocation.

This has been broken since January 2014, from a change made during
the scheduler refactor (527da4ca8a).

How nobody seems to have figured this out since then, I have no idea,
especially since the 32-bit initialization routine had the
critical function (x86_get_double_fault_stack) commented out entirely.

Even if it wasn't commented out (and it wasn't for 64-bit), it wouldn't
have worked, because the double-fault stacks were not allocated
until "post-VM", while the TSS is set up during "preboot" (basically
just after kernel entry.)

We now take care of this by allocating the virtual address for
the stacks immediately, and then creating an area for it
in the "post-VM" stage. This means that any double-faults
which occur before the post-VM stage will just triple-fault,
but seeing as all double-faults have been triple-faults
for the past decade, that's probably fine. (In order to get a
double-fault stack that early, we would probably need to
have the bootloader allocate it for us.)

This must be working, because it changes #18692
from a triple-fault (instant reboot) into a double-fault KDL.

src/system/kernel/arch/x86/32/descriptors.cpp

@@ -211,8 +211,8 @@ init_double_fault(int cpuNum)
 	struct tss* tss = &gCPU[cpuNum].arch.double_fault_tss;
 
 	memset(tss, 0, sizeof(struct tss));
-	size_t stackSize = 0;
-	//tss->sp0 = (addr_t)x86_get_double_fault_stack(cpuNum, &stackSize);
+	size_t stackSize;
+	tss->sp0 = (addr_t)x86_get_double_fault_stack(cpuNum, &stackSize);
 	tss->sp0 += stackSize;
 	tss->ss0 = KERNEL_DATA_SELECTOR;
 	tss->cr3 = x86_read_cr3();

src/system/kernel/arch/x86/arch_cpu.cpp

@@ -109,7 +109,7 @@ static uint32 sCpuRendezvous3;
 static vint32 sTSCSyncRendezvous;
 
 /* Some specials for the double fault handler */
-static uint8* sDoubleFaultStacks;
+static addr_t sDoubleFaultStacks = 0;
 static const size_t kDoubleFaultStackSize = 4096;	// size per CPU
 
 static x86_cpu_module_info* sCpuModule;
@@ -1480,7 +1480,7 @@ void*
 x86_get_double_fault_stack(int32 cpu, size_t* _size)
 {
 	*_size = kDoubleFaultStackSize;
-	return sDoubleFaultStacks + kDoubleFaultStackSize * cpu;
+	return (void*)(sDoubleFaultStacks + kDoubleFaultStackSize * cpu);
 }
 
 
@@ -1491,7 +1491,7 @@ int32
 x86_double_fault_get_cpu(void)
 {
 	addr_t stack = x86_get_stack_frame();
-	return (stack - (addr_t)sDoubleFaultStacks) / kDoubleFaultStackSize;
+	return (stack - sDoubleFaultStacks) / kDoubleFaultStackSize;
 }
 
 
@@ -1501,6 +1501,12 @@ x86_double_fault_get_cpu(void)
 status_t
 arch_cpu_preboot_init_percpu(kernel_args* args, int cpu)
 {
+	if (cpu == 0) {
+		// We can't allocate pages at this stage in the boot process, only virtual addresses.
+		sDoubleFaultStacks = vm_allocate_early(args,
+			kDoubleFaultStackSize * smp_get_num_cpus(), 0, 0, 0);
+	}
+
 	// On SMP system we want to synchronize the CPUs' TSCs, so system_time()
 	// will return consistent values.
 	if (smp_get_num_cpus() > 1) {
@@ -1774,24 +1780,20 @@ enable_xsavemask(void* dummy, int cpu)
 status_t
 arch_cpu_init_post_vm(kernel_args* args)
 {
-	uint32 i;
-
-	// allocate an area for the double fault stacks
-	virtual_address_restrictions virtualRestrictions = {};
-	virtualRestrictions.address_specification = B_ANY_KERNEL_ADDRESS;
-	physical_address_restrictions physicalRestrictions = {};
-	create_area_etc(B_SYSTEM_TEAM, "double fault stacks",
-		kDoubleFaultStackSize * smp_get_num_cpus(), B_FULL_LOCK,
-		B_KERNEL_READ_AREA | B_KERNEL_WRITE_AREA, CREATE_AREA_DONT_WAIT, 0,
-		&virtualRestrictions, &physicalRestrictions,
-		(void**)&sDoubleFaultStacks);
+	// allocate the area for the double fault stacks
+	area_id stacks = create_area("double fault stacks",
+		(void**)&sDoubleFaultStacks, B_EXACT_ADDRESS,
+		kDoubleFaultStackSize * smp_get_num_cpus(),
+		B_FULL_LOCK, B_KERNEL_READ_AREA | B_KERNEL_WRITE_AREA);
+	if (stacks < B_OK)
+		panic("failed to create double fault stacks area: %d", stacks);
 
 	X86PagingStructures* kernelPagingStructures
 		= static_cast<X86VMTranslationMap*>(
 			VMAddressSpace::Kernel()->TranslationMap())->PagingStructures();
 
 	// Set active translation map on each CPU.
-	for (i = 0; i < args->num_cpus; i++) {
+	for (uint32 i = 0; i < args->num_cpus; i++) {
 		gCPU[i].arch.active_paging_structures = kernelPagingStructures;
 		kernelPagingStructures->AddReference();
 	}
@@ -1847,7 +1849,6 @@ arch_cpu_init_post_vm(kernel_args* args)
 		dprintf("enable %s 0x%" B_PRIx64 " %" B_PRId64 "\n",
 			gHasXsavec ? "XSAVEC" : "XSAVE", gXsaveMask, gFPUSaveLength);
 	}
-
 #endif
 
 	return B_OK;