x86, amd: Avoid cache aliasing penalties on AMD family 15h

This patch provides performance tuning for the "Bulldozer" CPU. With its
shared instruction cache there is a chance of generating an excessive
number of cache cross-invalidates when running specific workloads on the
cores of a compute module.

This excessive amount of cross-invalidations can be observed if cache
lines backed by shared physical memory alias in bits [14:12] of their
virtual addresses, as those bits are used for the index generation.

This patch addresses the issue by clearing all the bits in the [14:12]
slice of the file mapping's virtual address at generation time, thus
forcing those bits the same for all mappings of a single shared library
across processes and, in doing so, avoids instruction cache aliases.

It also adds the command line option "align_va_addr=(32|64|on|off)" with
which virtual address alignment can be enabled for 32-bit or 64-bit x86
individually, or both, or be completely disabled.

This change leaves virtual region address allocation on other families
and/or vendors unaffected.

Signed-off-by: Borislav Petkov <[e-mail]>
Link: http://lkml.kernel.org/r/[e-mail]
Signed-off-by: H. Peter Anvin <[e-mail]>

Documentation/kernel-parameters.txt

···
 			behaviour to be specified.  Bit 0 enables warnings,
 			bit 1 enables fixups, and bit 2 sends a segfault.
 
+	align_va_addr=	[X86-64]
+			Align virtual addresses by clearing slice [14:12] when
+			allocating a VMA at process creation time. This option
+			gives you up to 3% performance improvement on AMD F15h
+			machines (where it is enabled by default) for a
+			CPU-intensive style benchmark, and it can vary highly in
+			a microbenchmark depending on workload and compiler.
+
+			1: only for 32-bit processes
+			2: only for 64-bit processes
+			on: enable for both 32- and 64-bit processes
+			off: disable for both 32- and 64-bit processes
+
 	amd_iommu=	[HW,X86-84]
 			Pass parameters to the AMD IOMMU driver in the system.
 			Possible values are:

arch/x86/include/asm/elf.h

···
 /*
  * ELF register definitions..
  */
+#include <linux/thread_info.h>
 
 #include <asm/ptrace.h>
 #include <asm/user.h>
···
 extern unsigned long arch_randomize_brk(struct mm_struct *mm);
 #define arch_randomize_brk arch_randomize_brk
 
+/*
+ * True on X86_32 or when emulating IA32 on X86_64
+ */
+static inline int mmap_is_ia32(void)
+{
+#ifdef CONFIG_X86_32
+	return 1;
+#endif
+#ifdef CONFIG_IA32_EMULATION
+	if (test_thread_flag(TIF_IA32))
+		return 1;
+#endif
+	return 0;
+}
+
+/* The first two values are special, do not change. See align_addr() */
+enum align_flags {
+	ALIGN_VA_32	= BIT(0),
+	ALIGN_VA_64	= BIT(1),
+	ALIGN_VDSO	= BIT(2),
+	ALIGN_TOPDOWN	= BIT(3),
+};
+
+struct va_alignment {
+	int flags;
+	unsigned long mask;
+} ____cacheline_aligned;
+
+extern struct va_alignment va_align;
+extern unsigned long align_addr(unsigned long, struct file *, enum align_flags);
 #endif /* _ASM_X86_ELF_H */

arch/x86/kernel/cpu/amd.c

···
 					"with P0 frequency!\n");
 		}
 	}
+
+	if (c->x86 == 0x15) {
+		unsigned long upperbit;
+		u32 cpuid, assoc;
+
+		cpuid	 = cpuid_edx(0x80000005);
+		assoc	 = cpuid >> 16 & 0xff;
+		upperbit = ((cpuid >> 24) << 10) / assoc;
+
+		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
+		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
+
+	}
 }
 
 static void __cpuinit init_amd(struct cpuinfo_x86 *c)

arch/x86/kernel/sys_x86_64.c

···
 #include <asm/ia32.h>
 #include <asm/syscalls.h>
 
+struct __read_mostly va_alignment va_align = {
+	.flags = -1,
+};
+
+/*
+ * Align a virtual address to avoid aliasing in the I$ on AMD F15h.
+ *
+ * @flags denotes the allocation direction - bottomup or topdown -
+ * or vDSO; see call sites below.
+ */
+unsigned long align_addr(unsigned long addr, struct file *filp,
+			 enum align_flags flags)
+{
+	unsigned long tmp_addr;
+
+	/* handle 32- and 64-bit case with a single conditional */
+	if (va_align.flags < 0 || !(va_align.flags & (2 - mmap_is_ia32())))
+		return addr;
+
+	if (!(current->flags & PF_RANDOMIZE))
+		return addr;
+
+	if (!((flags & ALIGN_VDSO) || filp))
+		return addr;
+
+	tmp_addr = addr;
+
+	/*
+	 * We need an address which is <= than the original
+	 * one only when in topdown direction.
+	 */
+	if (!(flags & ALIGN_TOPDOWN))
+		tmp_addr += va_align.mask;
+
+	tmp_addr &= ~va_align.mask;
+
+	return tmp_addr;
+}
+
+static int __init control_va_addr_alignment(char *str)
+{
+	/* guard against enabling this on other CPU families */
+	if (va_align.flags < 0)
+		return 1;
+
+	if (*str == 0)
+		return 1;
+
+	if (*str == '=')
+		str++;
+
+	if (!strcmp(str, "32"))
+		va_align.flags = ALIGN_VA_32;
+	else if (!strcmp(str, "64"))
+		va_align.flags = ALIGN_VA_64;
+	else if (!strcmp(str, "off"))
+		va_align.flags = 0;
+	else if (!strcmp(str, "on"))
+		va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
+	else
+		return 0;
+
+	return 1;
+}
+__setup("align_va_addr", control_va_addr_alignment);
+
 SYSCALL_DEFINE6(mmap, unsigned long, addr, unsigned long, len,
 		unsigned long, prot, unsigned long, flags,
 		unsigned long, fd, unsigned long, off)
···
 	start_addr = addr;
 
 full_search:
+
+	addr = align_addr(addr, filp, 0);
+
 	for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 		/* At this point:  (!vma || addr < vma->vm_end). */
 		if (end - len < addr) {
···
 			mm->cached_hole_size = vma->vm_start - addr;
 
 		addr = vma->vm_end;
+		addr = align_addr(addr, filp, 0);
 	}
 }
 
···
 
 	/* make sure it can fit in the remaining address space */
 	if (addr > len) {
-		vma = find_vma(mm, addr-len);
-		if (!vma || addr <= vma->vm_start)
+		unsigned long tmp_addr = align_addr(addr - len, filp,
+						    ALIGN_TOPDOWN);
+
+		vma = find_vma(mm, tmp_addr);
+		if (!vma || tmp_addr + len <= vma->vm_start)
 			/* remember the address as a hint for next time */
-			return mm->free_area_cache = addr-len;
+			return mm->free_area_cache = tmp_addr;
 	}
 
 	if (mm->mmap_base < len)
···
 	addr = mm->mmap_base-len;
 
 	do {
+		addr = align_addr(addr, filp, ALIGN_TOPDOWN);
+
 		/*
 		 * Lookup failure means no vma is above this address,
 		 * else if new region fits below vma->vm_start,

arch/x86/mm/mmap.c

···
 #define MIN_GAP (128*1024*1024UL + stack_maxrandom_size())
 #define MAX_GAP (TASK_SIZE/6*5)
 
-/*
- * True on X86_32 or when emulating IA32 on X86_64
- */
-static int mmap_is_ia32(void)
-{
-#ifdef CONFIG_X86_32
-	return 1;
-#endif
-#ifdef CONFIG_IA32_EMULATION
-	if (test_thread_flag(TIF_IA32))
-		return 1;
-#endif
-	return 0;
-}
-
 static int mmap_is_legacy(void)
 {
 	if (current->personality & ADDR_COMPAT_LAYOUT)

arch/x86/vdso/vma.c

···
 	addr = start + (offset << PAGE_SHIFT);
 	if (addr >= end)
 		addr = end;
+
+	/*
+	 * page-align it here so that get_unmapped_area doesn't
+	 * align it wrongfully again to the next page. addr can come in 4K
+	 * unaligned here as a result of stack start randomization.
+	 */
+	addr = PAGE_ALIGN(addr);
+	addr = align_addr(addr, NULL, ALIGN_VDSO);
+
 	return addr;
 }