Introduce and use a per-CPU read-mostly sub-section

Since mixing data that only gets setup once and then (perhaps
frequently) gets read by remote CPUs with data that the local CPU may
modify (again, perhaps frequently) still causes undesirable cache
protocol related bus traffic, separate the former class of objects
from the latter.

These objects converted here are just picked based on their write-once
(or write-very-rarely) properties; perhaps some more adjustments may
be desirable subsequently. The primary users of the new sub-section
will result from the next patch.

Signed-off-by: Jan Beulich <jbeulich@novell.com>

xen/arch/ia64/xen/vhpt.c

@@ -21,8 +21,8 @@
 #include <asm/vcpumask.h>
 #include <asm/vmmu.h>
 
-DEFINE_PER_CPU (unsigned long, vhpt_paddr);
-DEFINE_PER_CPU (unsigned long, vhpt_pend);
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, vhpt_paddr);
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, vhpt_pend);
 #ifdef CONFIG_XEN_IA64_TLBFLUSH_CLOCK
 DEFINE_PER_CPU(volatile u32, vhpt_tlbflush_timestamp);
 #endif

xen/arch/ia64/xen/xen.lds.S

@@ -187,6 +187,8 @@ SECTIONS
 	{
 		__per_cpu_start = .;
 		*(.data.percpu)
+		. = ALIGN(SMP_CACHE_BYTES);
+		*(.data.percpu.read_mostly)
 		__per_cpu_end = .;
 	}
   . = __phys_per_cpu_start + PERCPU_PAGE_SIZE;	/* ensure percpu data fits

xen/arch/x86/hvm/vmx/vmcs.c

@@ -66,7 +66,7 @@ u32 vmx_vmexit_control __read_mostly;
 u32 vmx_vmentry_control __read_mostly;
 bool_t cpu_has_vmx_ins_outs_instr_info __read_mostly;
 
-static DEFINE_PER_CPU(struct vmcs_struct *, host_vmcs);
+static DEFINE_PER_CPU_READ_MOSTLY(struct vmcs_struct *, host_vmcs);
 static DEFINE_PER_CPU(struct vmcs_struct *, current_vmcs);
 static DEFINE_PER_CPU(struct list_head, active_vmcs_list);
 

xen/arch/x86/setup.c

@@ -111,9 +111,9 @@ extern void vesa_init(void);
 extern void vesa_mtrr_init(void);
 extern void init_tmem(void);
 
-DEFINE_PER_CPU(struct desc_struct *, gdt_table) = boot_cpu_gdt_table;
+DEFINE_PER_CPU_READ_MOSTLY(struct desc_struct *, gdt_table) = boot_cpu_gdt_table;
 #ifdef CONFIG_COMPAT
-DEFINE_PER_CPU(struct desc_struct *, compat_gdt_table)
+DEFINE_PER_CPU_READ_MOSTLY(struct desc_struct *, compat_gdt_table)
     = boot_cpu_compat_gdt_table;
 #endif
 

xen/arch/x86/traps.c

@@ -78,7 +78,7 @@ char opt_nmi[10] = "fatal";
 #endif
 string_param("nmi", opt_nmi);
 
-DEFINE_PER_CPU(u32, ler_msr);
+DEFINE_PER_CPU_READ_MOSTLY(u32, ler_msr);
 
 /* Master table, used by CPU0. */
 idt_entry_t idt_table[IDT_ENTRIES];

xen/arch/x86/x86_32/traps.c

@@ -188,7 +188,7 @@ void show_page_walk(unsigned long addr)
     unmap_domain_page(l1t);
 }
 
-DEFINE_PER_CPU(struct tss_struct *, doublefault_tss);
+DEFINE_PER_CPU_READ_MOSTLY(struct tss_struct *, doublefault_tss);
 static unsigned char __attribute__ ((__section__ (".bss.page_aligned")))
     boot_cpu_doublefault_space[PAGE_SIZE];
 

xen/arch/x86/x86_32/xen.lds.S

@@ -4,6 +4,7 @@
  */
 
 #include <xen/config.h>
+#include <xen/cache.h>
 #include <asm/page.h>
 #include <asm/percpu.h>
 #undef ENTRY
@@ -69,9 +70,13 @@ SECTIONS
   __init_end = .;
 
   __per_cpu_shift = PERCPU_SHIFT; /* kdump assist */
-  __per_cpu_start = .;
-  .data.percpu : { *(.data.percpu) } :text
-  __per_cpu_data_end = .;
+  .data.percpu : {
+	__per_cpu_start = .;
+	*(.data.percpu)
+	. = ALIGN(SMP_CACHE_BYTES);
+	*(.data.percpu.read_mostly)
+	__per_cpu_data_end = .;
+  } :text
   . = __per_cpu_start + (NR_CPUS << PERCPU_SHIFT);
   . = ALIGN(PAGE_SIZE);
   __per_cpu_end = .;

xen/arch/x86/x86_64/mm.c

@@ -37,7 +37,7 @@
 unsigned int m2p_compat_vstart = __HYPERVISOR_COMPAT_VIRT_START;
 #endif
 
-DEFINE_PER_CPU(void *, compat_arg_xlat);
+DEFINE_PER_CPU_READ_MOSTLY(void *, compat_arg_xlat);
 
 /* Top-level master (and idle-domain) page directory. */
 l4_pgentry_t __attribute__ ((__section__ (".bss.page_aligned")))

xen/arch/x86/x86_64/xen.lds.S

@@ -2,6 +2,7 @@
 /* Modified for x86-64 Xen by Keir Fraser */
 
 #include <xen/config.h>
+#include <xen/cache.h>
 #include <asm/page.h>
 #include <asm/percpu.h>
 #undef ENTRY
@@ -67,9 +68,13 @@ SECTIONS
   __init_end = .;
 
   __per_cpu_shift = PERCPU_SHIFT; /* kdump assist */
-  __per_cpu_start = .;
-  .data.percpu : { *(.data.percpu) } :text
-  __per_cpu_data_end = .;
+  .data.percpu : {
+	__per_cpu_start = .;
+	*(.data.percpu)
+	. = ALIGN(SMP_CACHE_BYTES);
+	*(.data.percpu.read_mostly)
+	__per_cpu_data_end = .;
+  } :text
   . = __per_cpu_start + (NR_CPUS << PERCPU_SHIFT);
   . = ALIGN(PAGE_SIZE);
   __per_cpu_end = .;

xen/common/kexec.c

@@ -27,7 +27,7 @@
 #include <compat/kexec.h>
 #endif
 
-static DEFINE_PER_CPU(void *, crash_notes);
+static DEFINE_PER_CPU_READ_MOSTLY(void *, crash_notes);
 
 static Elf_Note *xen_crash_note;
 

xen/common/tmem_xen.c

@@ -36,8 +36,8 @@ DECL_CYC_COUNTER(pg_copy);
  * allocated iff opt_tmem_compress */
 #define LZO_WORKMEM_BYTES LZO1X_1_MEM_COMPRESS
 #define LZO_DSTMEM_PAGES 2
-static DEFINE_PER_CPU(unsigned char *, workmem);
-static DEFINE_PER_CPU(unsigned char *, dstmem);
+static DEFINE_PER_CPU_READ_MOSTLY(unsigned char *, workmem);
+static DEFINE_PER_CPU_READ_MOSTLY(unsigned char *, dstmem);
 
 #ifdef COMPARE_COPY_PAGE_SSE2
 #include <asm/flushtlb.h>  /* REMOVE ME AFTER TEST */

xen/common/trace.c

@@ -46,8 +46,8 @@ static unsigned int opt_tbuf_size = 0;
 integer_param("tbuf_size", opt_tbuf_size);
 
 /* Pointers to the meta-data objects for all system trace buffers */
-static DEFINE_PER_CPU(struct t_buf *, t_bufs);
-static DEFINE_PER_CPU(unsigned char *, t_data);
+static DEFINE_PER_CPU_READ_MOSTLY(struct t_buf *, t_bufs);
+static DEFINE_PER_CPU_READ_MOSTLY(unsigned char *, t_data);
 static int data_size;
 
 /* High water mark for trace buffers; */

xen/include/asm-ia64/linux-xen/asm/percpu.h

@@ -26,9 +26,9 @@
 	extern __SMALL_ADDR_AREA __typeof__(type) per_cpu__##name
 
 /* Separate out the type, so (int[3], foo) works. */
-#define DEFINE_PER_CPU(type, name)				\
-	__attribute__((__section__(".data.percpu")))		\
-	__SMALL_ADDR_AREA __typeof__(type) per_cpu__##name
+#define __DEFINE_PER_CPU(type, name, suffix)				\
+	__attribute__((__section__(".data.percpu" #suffix)))		\
+	__SMALL_ADDR_AREA __typeof__(type) per_cpu_##name
 
 /*
  * Pretty much a literal copy of asm-generic/percpu.h, except that percpu_modcopy() is an

xen/include/asm-x86/percpu.h

@@ -5,9 +5,9 @@
 #define PERCPU_SIZE  (1UL << PERCPU_SHIFT)
 
 /* Separate out the type, so (int[3], foo) works. */
-#define DEFINE_PER_CPU(type, name)                      \
-    __attribute__((__section__(".data.percpu")))        \
-    __typeof__(type) per_cpu__##name
+#define __DEFINE_PER_CPU(type, name, suffix)                    \
+    __attribute__((__section__(".data.percpu" #suffix)))        \
+    __typeof__(type) per_cpu_##name
 
 /* var is in discarded region: offset to particular copy we want */
 #define per_cpu(var, cpu)  \

xen/include/xen/percpu.h

@@ -4,6 +4,16 @@
 #include <xen/config.h>
 #include <asm/percpu.h>
 
+/*
+ * Separate out the type, so (int[3], foo) works.
+ *
+ * The _##name concatenation is being used here to prevent 'name' from getting
+ * macro expanded, while still allowing a per-architecture symbol name prefix.
+ */
+#define DEFINE_PER_CPU(type, name) __DEFINE_PER_CPU(type, _##name, )
+#define DEFINE_PER_CPU_READ_MOSTLY(type, name) \
+	__DEFINE_PER_CPU(type, _##name, .read_mostly)
+
 /* Preferred on Xen. Also see arch-defined per_cpu(). */
 #define this_cpu(var)    __get_cpu_var(var)