multikernel: support isolated SR-IOV VF assignment

arch/x86/include/asm/multikernel.h

@@ -10,6 +10,7 @@
 
 #ifndef __ASSEMBLY__
 
+#include <linux/init.h>
 #include <linux/types.h>
 #include <linux/cpumask.h>
 #include <linux/sizes.h>
@@ -54,17 +55,21 @@ static inline int arch_cpu_from_physical_id(u64 phys_id)
  * from a page written by the host while running on CPUs parked by (possibly
  * differently built) spawn kernels.
  *
- * The fields fall into two classes that must not be mixed up:
+ * The fields fall into three classes that must not be mixed up:
  *
  *  - Anchor fields (self_phys, park_phys, park_cr3, ctrl_phys,
  *    ctrl_size): the context's own identity, written once when the
  *    context is set up. The spawn kernel reads them whenever it parks a
  *    CPU on halt or offline, so they must stay valid for the context's
  *    whole lifetime.
  *
- *  - Dispatch fields (everything else): the wake mailbox, rewritten for
- *    every publication and staged into registers by the CPU that claims
- *    it. Reparking gets its own repark_* dispatch fields precisely so a
+ *  - Primary boot fields (boot_* and bp): written before the boot CPU is
+ *    released and consumed while that kernel initializes. Secondary and
+ *    repark publications do not rewrite them.
+ *
+ *  - Dispatch fields (the remaining fixed-size fields): the wake mailbox,
+ *    rewritten for every publication and staged into registers by the CPU
+ *    that claims it. Reparking gets its own repark_* dispatch fields so a
  *    repark publication never overwrites the anchor: the two used to
  *    share fields, and a repark left the anchor pointing at another
  *    kernel's park area, which triple-faulted the next halt.
@@ -80,6 +85,10 @@ struct mk_spawn_context {
 	unsigned long gs_base;		/* Per-CPU GS base (for secondary) */
 	unsigned long stack;		/* Stack pointer (for secondary) */
 	unsigned long spawn_cr3;	/* Spawn kernel's CR3 (for secondary CPU final switch) */
+	unsigned long boot_lps;		/* Host delay loops per second */
+	unsigned long boot_cpu_khz;	/* Host CPU frequency calibration */
+	unsigned long boot_tsc_khz;	/* Host TSC frequency calibration */
+	unsigned long boot_apic_hz;	/* Host local APIC timer frequency */
 	unsigned long park_phys;	/* Pool park code page (host owned, never reloaded) */
 	unsigned long park_cr3;		/* Page table parked CPUs run on (host owned) */
 	unsigned long repark_park_phys;	/* REPARK dispatch: park page to move to */
@@ -161,6 +170,19 @@ int multikernel_wakeup_secondary_cpu_64(u32 apicid, unsigned long start_eip,
 int multikernel_restore_ap(unsigned int cpu, unsigned long cr3,
 			   unsigned long gs_base, unsigned long stack,
 			   unsigned long entry);
+struct mk_pci_host_bridge;
+
+#ifdef CONFIG_MULTIKERNEL
+int mk_arch_snapshot_pci_host_bridges(const struct mk_instance *instance,
+				      struct mk_pci_host_bridge *bridges,
+				      size_t capacity);
+#endif
+
+#if defined(CONFIG_MULTIKERNEL) && defined(CONFIG_PCI)
+void __init x86_multikernel_pci_platform_init(void);
+#else
+static inline void x86_multikernel_pci_platform_init(void) { }
+#endif
 
 #endif /* __ASSEMBLY__ */
 

arch/x86/include/asm/pci_x86.h

@@ -190,6 +190,10 @@ extern struct pci_mmcfg_region *__init pci_mmconfig_add(int segment, int start,
 
 extern struct list_head pci_mmcfg_list;
 
+typedef int (*pci_mmcfg_region_cb)(const struct pci_mmcfg_region *region,
+				   void *data);
+int pci_mmcfg_walk_regions(pci_mmcfg_region_cb callback, void *data);
+
 #define PCI_MMCFG_BUS_OFFSET(bus)      ((bus) << 20)
 
 /*

arch/x86/kernel/platform-quirks.c

@@ -14,8 +14,10 @@
 #include <asm/e820/api.h>
 #include <asm/apic.h>
 #include <asm/apicdef.h>
+#include <asm/cpufeature.h>
 #include <asm/mpspec.h>
 #include <asm/numa.h>
+#include <asm/page.h>
 #include <linux/pgtable.h>
 #include <asm/pgtable.h>
 #include <asm/pgtable_64_types.h>
@@ -28,6 +30,35 @@ extern pmd_t *populate_extra_pmd(unsigned long vaddr);
 extern unsigned long orig_boot_params;
 
 #ifdef CONFIG_MULTIKERNEL
+static unsigned long multikernel_cpu_khz;
+static unsigned long multikernel_tsc_khz;
+
+static unsigned long multikernel_calibrate_cpu(void)
+{
+	return multikernel_cpu_khz;
+}
+
+static unsigned long multikernel_calibrate_tsc(void)
+{
+	return multikernel_tsc_khz;
+}
+
+static void __init multikernel_setup_calibration(void)
+{
+	phys_addr_t ctx_phys = orig_boot_params -
+		offsetof(struct mk_spawn_context, bp);
+	struct mk_spawn_context *ctx = __va(ctx_phys);
+
+	if (ctx->self_phys != ctx_phys || !ctx->boot_tsc_khz)
+		return;
+
+	multikernel_tsc_khz = ctx->boot_tsc_khz;
+	multikernel_cpu_khz = ctx->boot_cpu_khz ?: ctx->boot_tsc_khz;
+	x86_platform.calibrate_cpu = multikernel_calibrate_cpu;
+	x86_platform.calibrate_tsc = multikernel_calibrate_tsc;
+	setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
+}
+
 /*
  * Custom wakeup for multikernel spawn kernels.
  * Uses shared spawn table instead of realmode trampoline.
@@ -105,6 +136,10 @@ static void __init multikernel_parse_smp_config(void)
 	 */
 	apic_update_callback(wakeup_secondary_cpu_64, multikernel_wakeup_cpu);
 }
+#else
+static inline void multikernel_setup_calibration(void)
+{
+}
 #endif /* CONFIG_MULTIKERNEL */
 
 void __init x86_early_init_platform_quirks(void)
@@ -135,6 +170,8 @@ void __init x86_early_init_platform_quirks(void)
 		x86_platform.legacy.i8042 = X86_LEGACY_I8042_PLATFORM_ABSENT;
 		break;
 	case X86_SUBARCH_MULTIKERNEL:
+		multikernel_setup_calibration();
+		x86_multikernel_pci_platform_init();
 		x86_platform.legacy.devices.pnpbios = 0;
 		x86_platform.legacy.i8042 = X86_LEGACY_I8042_PLATFORM_ABSENT;
 		x86_platform.legacy.rtc = 0;
@@ -175,7 +212,9 @@ void __init x86_early_init_platform_quirks(void)
 		 * the PIT - which belongs to the host - and then request
 		 * legacy IRQ0, which can never reach an instance CPU that
 		 * has neither a PIC nor an IO-APIC. Ticks come from the
-		 * local APIC timer via setup_percpu_clockev() instead.
+		 * local APIC timer initialized by setup_percpu_clockev().
+		 * Keeping global_clock_event unset bypasses LAPIC timer
+		 * verification, whose fallback path requires legacy IRQ0.
 		 */
 		x86_init.timers.timer_init = x86_init_noop;
 		x86_init.timers.wallclock_init = x86_init_noop;

arch/x86/multikernel/Makefile

@@ -4,3 +4,4 @@
 #
 
 obj-y += spawn.o direct_boot.o head_64.o
+obj-$(CONFIG_PCI) += pci.o