ARM: kernel: update cpuinfo to print all online CPUs features

RFC PATCH New Lorenzo Pieralisi 3 participants 4 messages Oct 16, 2012 – Oct 17, 2012 Arm (32-bit)

Share & tools

Lorenzo Pieralisi

author

2012-10-16 16:24 UTC #0 lore raw

RFC PATCH 1/1 ARM: kernel: update cpuinfo to print all online CPUs features

Currently, reading /proc/cpuinfo provides userspace with CPU ID of the CPU carrying out the read from the file. This is fine as long as all CPUs in the system are the same. With the advent of big.LITTLE and heterogenous ARM systems this approach provides user space with incorrect bits of information since CPU ids in the system might differ from the one provided by the CPU reading the file.

This patch updates the cpuinfo show function and some internal data structures so that a read from /proc/cpuinfo prints HW information for all online CPUs at once, mirroring x86 behaviour.

Signed-off-by Lorenzo Pieralisi <[email protected]>

---

Posting it as a way to get advice on the best way to improve /proc/cpuinfo for heterogenous multi-cluster systems and to provide proper information to userspace without breaking existing semantics, no more than that.

Comments more than welcome.

Thanks,
Lorenzo

 arch/arm/include/asm/cpu.h |  1 +
 arch/arm/kernel/setup.c    | 66 +++++++++++++++++++++++-----------------------
 arch/arm/kernel/smp.c      |  1 +
 3 files changed, 35 insertions(+), 33 deletions(-)

arch/arm/include/asm/cpu.h +1 -0

index d797223..2744f06 100644
@@ -15,6 +15,7 @@
 
 struct cpuinfo_arm {
 	struct cpu	cpu;
+	u32		cpuid;
 #ifdef CONFIG_SMP
 	unsigned int	loops_per_jiffy;
 #endif

arch/arm/kernel/setup.c +33 ~~-33~~

88 lines not on this page · raw message

arch/arm/kernel/smp.c +1 -0

index f0bcb1e..73f407e 100644
@@ -211,6 +211,7 @@ static void __cpuinit smp_store_cpu_info(unsigned int cpuid)
 	struct cpuinfo_arm *cpu_info = &per_cpu(cpu_data, cpuid);
 
 	cpu_info->loops_per_jiffy = loops_per_jiffy;
+	cpu_info->cpuid = read_cpuid_id();
 
 	store_cpu_topology(cpuid);
 }

Signature · 1 line

1.7.12

Santosh Shilimkar

2012-10-16 17:59 UTC #1 lore raw

RFC PATCH 1/1 ARM: kernel: update cpuinfo to print all online CPUs features

Quoted 28 lines · Lorenzo Pieralisi

Currently, reading /proc/cpuinfo provides userspace with CPU ID of the CPU carrying out the read from the file. This is fine as long as all CPUs in the system are the same. With the advent of big.LITTLE and heterogenous ARM systems this approach provides user space with incorrect bits of information since CPU ids in the system might differ from the one provided by the CPU reading the file.
This patch updates the cpuinfo show function and some internal data structures so that a read from /proc/cpuinfo prints HW information for all online CPUs at once, mirroring x86 behaviour.
Signed-off-by Lorenzo Pieralisi <[email protected]>
---
Posting it as a way to get advice on the best way to improve /proc/cpuinfo for heterogenous multi-cluster systems and to provide proper information to userspace without breaking existing semantics, no more than that.
Comments more than welcome.
Thanks,
Lorenzo
  arch/arm/include/asm/cpu.h |  1 +
  arch/arm/kernel/setup.c    | 66 +++++++++++++++++++++++-----------------------
  arch/arm/kernel/smp.c      |  1 +
  3 files changed, 35 insertions(+), 33 deletions(-)

[...]

Quoted 9 lines

arch/arm/kernel/setup.c +0 -1

index 8707cff..bf7839d08 100644
@@ -1054,13 +1054,15 @@ static const char *hwcap_str[] = {
 
  static int c_show(struct seq_file *m, void *v)
  {
-	int i;

arch/arm/kernel/setup.c static const char *hwcap_str[] = { · line 1064

Quoted 10 lines

arch/arm/kernel/setup.c +6 -2

+	int i, j;
+	u32 cpuid;
 
-	seq_printf(m, "Processor\t: %s rev %d (%s)\n",
-		   cpu_name, read_cpuid_id() & 15, elf_platform);
+	for_each_online_cpu(i) {
+		cpuid = is_smp() ? per_cpu(cpu_data, i).cpuid : read_cpuid_id();
+		seq_printf(m, "Processor\t: %s rev %d (%s)\n",
+			   cpu_name, cpuid & 15, elf_platform);

Not exactly related to the $subject patch, but I remember doing a patch to have cat /proc/cpuinfo spitting only online CPUs just like x86 using for_each_online_cpu(i).
At that point Russell mentioned about a possibility of read() syscall spreading over the hot-plug operation and hence the above may not be safe.

is that right Russell ?

regards
Santosh

Russell King - ARM Linux

2012-10-16 21:47 UTC #2 lore raw

RFC PATCH 1/1 ARM: kernel: update cpuinfo to print all online CPUs features

Quoted 8 lines · Santosh Shilimkar

Not exactly related to the $subject patch, but I remember doing a patch to have cat /proc/cpuinfo spitting only online CPUs just like x86 using for_each_online_cpu(i).
At that point Russell mentioned about a possibility of read() syscall spreading over the hot-plug operation and hence the above may not be safe.
is that right Russell ?

That is correct, but you will notice that the code now uses the online cpus rather than the present, inspite of my objections. I gave up fighting the case, so now people get to live with the consequences of this.

This will be fun with dynamic hotplugging with big.LITTLE when people come to read any of these files which change their output with the online CPUs. Oh what fun we're in for there.

I guess the only way to convince people is to let them make their mistakes in the kernel.

Lorenzo Pieralisi

author

2012-10-17 10:20 UTC #3 lore raw

RFC PATCH 1/1 ARM: kernel: update cpuinfo to print all online CPUs features

Quoted 10 lines · Russell King - ARM Linux

Quoted 8 lines · Santosh Shilimkar
Not exactly related to the $subject patch, but I remember doing a patch to have cat /proc/cpuinfo spitting only online CPUs just like x86 using for_each_online_cpu(i).
At that point Russell mentioned about a possibility of read() syscall spreading over the hot-plug operation and hence the above may not be safe.
is that right Russell ?

Quoted 8 lines

That is correct, but you will notice that the code now uses the online cpus rather than the present, inspite of my objections. I gave up fighting the case, so now people get to live with the consequences of this.
This will be fun with dynamic hotplugging with big.LITTLE when people come to read any of these files which change their output with the online CPUs. Oh what fun we're in for there.

I had a look at how glibc currently detects cpus in __get_nprocs() and all methods end up relying on online CPUs:

- /sys/devices/system/cpu/online
- /proc/stat
- /proc/cpuinfo

Given the thread:

http://lists.infradead.org/pipermail/linux-arm-kernel/2011-June/054081.html

I do not think there is room to change the current behaviour of /proc/cpuinfo as far as the online vs. present argument goes, and I certainly did not try.

I posted this code just to understand if it makes sense to expose all CPU ids in /proc/cpuinfo the way I did (the same way x86 does, even though the code is slightly different).

Quoted 2 lines

I guess the only way to convince people is to let them make their mistakes in the kernel.

That's what I wanted to prevent by posting this code, namely getting feedback on the best way to improve /proc/cpuinfo for systems where CPU ids are not homogeneous.

The "online vs. present" issue is already there as Russell pointed out.

Thank you very much for the review, any further feedback appreciated.

Lorenzo