cpu/x86/smm: Introduce SMM module loader version 2

Xeon-SP Skylake Scalable Processor can have 36 CPU threads (18 cores).
Current coreboot SMM is unable to handle more than ~32 CPU threads.
This patch introduces a version 2 of the SMM module loader which
addresses this problem. Having two versions of the SMM module loader
prevents any issues to current projects. Future Xeon-SP products will
be using this version of the SMM loader.  Subsequent patches will
enable board specific functionality for Xeon-SP.

The reason for moving to version 2 is the state save area begins to
encroach upon the SMI handling code when more than 32 CPU threads are
in the system. This can cause system hangs, reboots, etc. The second
change is related to staggered entry points with simple near jumps. In
the current loader, near jumps will not work because the CPU is jumping
within the same code segment. In version 2, "far" address jumps are
necessary therefore protected mode must be enabled first. The SMM
layout and how the CPUs are staggered are documented in the code.

By making the modifications above, this allows the smm module loader to
expand easily as more CPU threads are added.

TEST=build for Tiogapass platform under OCP mainboard. Enable the
following in Kconfig.
        select CPU_INTEL_COMMON_SMM
        select SOC_INTEL_COMMON_BLOCK_SMM
        select SMM_TSEG
        select HAVE_SMI_HANDLER
        select ACPI_INTEL_HARDWARE_SLEEP_VALUES

Debug console will show all 36 cores relocated. Further tested by
generating SMI's to port 0xb2 using XDP/ITP HW debugger and ensured all
cores entering and exiting SMM properly. In addition, booted to Linux
5.4 kernel and observed no issues during mp init.

Change-Id: I00a23a5f2a46110536c344254868390dbb71854c
Signed-off-by: Rocky Phagura <rphagura@fb.com>
Reviewed-on: https://review.coreboot.org/c/coreboot/+/43684
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Angel Pons <th3fanbus@gmail.com>

1 files changed, 31 insertions, 8 deletions

diff --git a/src/cpu/x86/mp_init.c b/src/cpu/x86/mp_init.c
index caed8f4005..5807831c98 100644
--- a/src/cpu/x86/mp_init.c
+++ b/src/cpu/x86/mp_init.c
@@ -726,12 +726,21 @@ static void asmlinkage smm_do_relocation(void *arg)
 	 * the location of the new SMBASE. If using SMM modules then this
 	 * calculation needs to match that of the module loader.
 	 */
+#if CONFIG(X86_SMM_LOADER_VERSION2)
+	perm_smbase = smm_get_cpu_smbase(cpu);
+	mp_state.perm_smbase = perm_smbase;
+	if (!perm_smbase) {
+		printk(BIOS_ERR, "%s: bad SMBASE for CPU %d\n", __func__, cpu);
+		return;
+	}
+#else
 	perm_smbase = mp_state.perm_smbase;
 	perm_smbase -= cpu * runtime->save_state_size;
-
-	printk(BIOS_DEBUG, "New SMBASE 0x%08lx\n", perm_smbase);
+#endif
 
 	/* Setup code checks this callback for validity. */
+	printk(BIOS_INFO, "%s : curr_smbase 0x%x perm_smbase 0x%x, cpu = %d\n",
+		__func__, (int)curr_smbase, (int)perm_smbase, cpu);
 	mp_state.ops.relocation_handler(cpu, curr_smbase, perm_smbase);
 
 	if (CONFIG(STM)) {
@@ -758,9 +767,17 @@ static void adjust_smm_apic_id_map(struct smm_loader_params *smm_params)
 
 static int install_relocation_handler(int num_cpus, size_t save_state_size)
 {
+	int cpus = num_cpus;
+#if CONFIG(X86_SMM_LOADER_VERSION2)
+	/* Default SMRAM size is not big enough to concurrently
+	 * handle relocation for more than ~32 CPU threads
+	 * therefore, relocate 1 by 1. */
+	cpus = 1;
+#endif
+
 	struct smm_loader_params smm_params = {
 		.per_cpu_stack_size = CONFIG_SMM_STUB_STACK_SIZE,
-		.num_concurrent_stacks = num_cpus,
+		.num_concurrent_stacks = cpus,
 		.per_cpu_save_state_size = save_state_size,
 		.num_concurrent_save_states = 1,
 		.handler = smm_do_relocation,
@@ -770,9 +787,10 @@ static int install_relocation_handler(int num_cpus, size_t save_state_size)
 	if (mp_state.ops.adjust_smm_params != NULL)
 		mp_state.ops.adjust_smm_params(&smm_params, 0);
 
-	if (smm_setup_relocation_handler(&smm_params))
+	if (smm_setup_relocation_handler(&smm_params)) {
+		printk(BIOS_ERR, "%s: smm setup failed\n", __func__);
 		return -1;
-
+	}
 	adjust_smm_apic_id_map(&smm_params);
 
 	return 0;
@@ -781,8 +799,13 @@ static int install_relocation_handler(int num_cpus, size_t save_state_size)
 static int install_permanent_handler(int num_cpus, uintptr_t smbase,
 					size_t smsize, size_t save_state_size)
 {
-	/* There are num_cpus concurrent stacks and num_cpus concurrent save
-	 * state areas. Lastly, set the stack size to 1KiB. */
+	/*
+	 * All the CPUs will relocate to permanaent handler now. Set parameters
+	 * needed for all CPUs. The placement of each CPUs entry point is
+	 * determined by the loader. This code simply provides the beginning of
+	 * SMRAM region, the number of CPUs who will use the handler, the stack
+	 * size and save state size for each CPU.
+	 */
 	struct smm_loader_params smm_params = {
 		.per_cpu_stack_size = CONFIG_SMM_MODULE_STACK_SIZE,
 		.num_concurrent_stacks = num_cpus,
@@ -794,7 +817,7 @@ static int install_permanent_handler(int num_cpus, uintptr_t smbase,
 	if (mp_state.ops.adjust_smm_params != NULL)
 		mp_state.ops.adjust_smm_params(&smm_params, 1);
 
-	printk(BIOS_DEBUG, "Installing SMM handler to 0x%08lx\n", smbase);
+	printk(BIOS_DEBUG, "Installing permanent SMM handler to 0x%08lx\n", smbase);
 
 	if (smm_load_module((void *)smbase, smsize, &smm_params))
 		return -1;