arm64: /proc/cpuinfo updates for compat tasks

The first patch is relatively simple, adding the "model name" line which is present in the arm32 /proc/cpuinfo.

The second patch is up for discussion. When arm64 support was merged, we concluded that "uname -m" should report "aarch64" even for compat tasks. The only way to change the output to "armv8l" was to set the PER_LINUX32 personality (and I think this is in line with x86).

We followed a similar personality approach for /proc/cpuinfo where the compat one is only exposed if the task sets/inherits PER_LINUX32. The downside is that apparently some 32-bit applications started making use of the 64-bit /proc/cpuinfo format. To prevent further spread, I'm proposing the second patch in this series to force compat /proc/cpuinfo for 32-bit tasks. This is a departure from the "uname -m" behaviour but maybe they don't need to be coupled anyway.

Thoughts?

Catalin Marinas (2):
arm64: Provide "model name" in /proc/cpuinfo for PER_LINUX32 tasks
arm64: Always provide compat /proc/cpuinfo for 32-bit tasks

 arch/arm64/include/asm/elf.h | 4 ++--
 arch/arm64/kernel/cpuinfo.c  | 9 ++++++++-
 2 files changed, 10 insertions(+), 3 deletions(-)

This patch brings the PER_LINUX32 /proc/cpuinfo format more in line with the 32-bit ARM one by providing an additional line:

model name : ARMv8 Processor rev X (v8l)

---
 arch/arm64/include/asm/elf.h | 4 ++--
 arch/arm64/kernel/cpuinfo.c  | 8 +++++++-
 2 files changed, 9 insertions(+), 3 deletions(-)

index 7a09c48c0475..579b6e654f2d 100644
@@ -160,14 +160,14 @@ extern int arch_setup_additional_pages(struct linux_binprm *bprm,
 #define STACK_RND_MASK			(0x3ffff >> (PAGE_SHIFT - 12))
 #endif
 
-#ifdef CONFIG_COMPAT
-
 #ifdef __AARCH64EB__
 #define COMPAT_ELF_PLATFORM		("v8b")
 #else
 #define COMPAT_ELF_PLATFORM		("v8l")
 #endif
 
+#ifdef CONFIG_COMPAT
+
 #define COMPAT_ELF_ET_DYN_BASE		(2 * TASK_SIZE_32 / 3)
 
 /* AArch32 registers. */

index 3808470486f3..c173d329397f 100644
@@ -22,6 +22,8 @@
 
 #include <linux/bitops.h>
 #include <linux/bug.h>
+#include <linux/compat.h>
+#include <linux/elf.h>
 #include <linux/init.h>
 #include <linux/kernel.h>
 #include <linux/personality.h>

@@ -104,6 +106,7 @@ static const char *const compat_hwcap2_str[] = {
 static int c_show(struct seq_file *m, void *v)
 {
 	int i, j;
+	bool compat = personality(current->personality) == PER_LINUX32;
 
 	for_each_online_cpu(i) {
 		struct cpuinfo_arm64 *cpuinfo = &per_cpu(cpu_data, i);

@@ -115,6 +118,9 @@ static int c_show(struct seq_file *m, void *v)
 		 * "processor".  Give glibc what it expects.
 		 */
 		seq_printf(m, "processor\t: %d\n", i);
+		if (compat)
+			seq_printf(m, "model name\t: ARMv8 Processor rev %d (%s)\n",
+				   MIDR_REVISION(midr), COMPAT_ELF_PLATFORM);
 
 		seq_printf(m, "BogoMIPS\t: %lu.%02lu\n",
 			   loops_per_jiffy / (500000UL/HZ),

@@ -127,7 +133,7 @@ static int c_show(struct seq_file *m, void *v)
 		 * software which does already (at least for 32-bit).
 		 */
 		seq_puts(m, "Features\t:");
-		if (personality(current->personality) == PER_LINUX32) {
+		if (compat) {
 #ifdef CONFIG_COMPAT
 			for (j = 0; compat_hwcap_str[j]; j++)
 				if (compat_elf_hwcap & (1 << j))

Looks good to me:

Will

I guess you can take it as a fix for 4.7.

Catalin

Currently, the compat /proc/cpuinfo is provided only if a task has the PER_LINUX32 personality, either by setting it explicitly or by inheriting it from the parent task. This is in line with the "uname -m" behaviour.

However, there are 32-bit user applications making use of the /proc/cpuinfo and unaware of a need to set the personality. This patch changes the arm64 /proc/cpuinfo logic to provide the compat information if the task is 32-bit _or_ the PER_LINUX32 personality is set.

---
 arch/arm64/kernel/cpuinfo.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

index c173d329397f..bd6a122ea856 100644
@@ -106,7 +106,8 @@ static const char *const compat_hwcap2_str[] = {
 static int c_show(struct seq_file *m, void *v)
 {
 	int i, j;
-	bool compat = personality(current->personality) == PER_LINUX32;
+	bool compat = is_compat_task() ||
+		personality(current->personality) == PER_LINUX32;
 
 	for_each_online_cpu(i) {
 		struct cpuinfo_arm64 *cpuinfo = &per_cpu(cpu_data, i);

Hi all,

What's the status on this patch? The previous patch (which was accepted already) is indeed confusing, because ARM32 processes on an ARM64 system are not necessarily setting PER_LINUX32.

I'm also curious, why was 'model name' removed from ARM64 in the first place? Plenty of other architectures support a similar property, and it's useful for some tools that already parse this, such as coreutils uname -p on Gentoo (and presumably others -- my Ubuntu machine must be similarly patched, as it supports uname -p on x86_64).

If we were to just stick 'model name' back in unconditionally, we could avoid the objections Will raised entirely :)

Regards,
Brian