Re: [PATCH v2] LoongArch: Correct the calculation logic of thread_count

From: Qiang Ma

Date: Thu Nov 27 2025 - 03:57:06 EST



在 2025/11/27 15:36, Huacai Chen 写道:
Hi, Qiang,

On Thu, Nov 27, 2025 at 11:32 AM Qiang Ma <maqianga@xxxxxxxxxxxxx> wrote:
For thread_count, the current calculation method has a maximum of 255,
which may not be sufficient in the future. Therefore, we are correcting
it now.

Reference: SMBIOS Specification, 7.5 Processor Information (Type 4)[1]

[1]: https://www.dmtf.org/sites/default/files/standards/documents/DSP0134_3.9.0.pdf

Signed-off-by: Qiang Ma <maqianga@xxxxxxxxxxxxx>
---
arch/loongarch/kernel/setup.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/arch/loongarch/kernel/setup.c b/arch/loongarch/kernel/setup.c
index 25a87378e48e..8dd25fc4bd13 100644
--- a/arch/loongarch/kernel/setup.c
+++ b/arch/loongarch/kernel/setup.c
@@ -56,6 +56,7 @@
#define SMBIOS_FREQLOW_MASK 0xFF
#define SMBIOS_CORE_PACKAGE_OFFSET 0x23
#define SMBIOS_THREAD_PACKAGE_OFFSET 0x25
+#define SMBIOS_THREAD_PACKAGE_2_OFFSET 0x2E
#define LOONGSON_EFI_ENABLE (1 << 3)

unsigned long fw_arg0, fw_arg1, fw_arg2;
@@ -120,13 +121,19 @@ static void __init parse_cpu_table(const struct dmi_header *dm)
{
long freq_temp = 0;
char *dmi_data = (char *)dm;
+ u8 thread_count;

freq_temp = ((*(dmi_data + SMBIOS_FREQHIGH_OFFSET) << 8) +
((*(dmi_data + SMBIOS_FREQLOW_OFFSET)) & SMBIOS_FREQLOW_MASK));
cpu_clock_freq = freq_temp * 1000000;

loongson_sysconf.cpuname = (void *)dmi_string_parse(dm, dmi_data[16]);
- loongson_sysconf.cores_per_package = *(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
+ thread_count = *(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
Yes, this one should need to be converted:
thread_count = *(u8 *)(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
+ if (thread_count != 0)
+ loongson_sysconf.cores_per_package =
+ dm->length >= 0x30 && thread_count == 0xFF ?
+ *(u16 *)(dmi_data + SMBIOS_THREAD_PACKAGE_2_OFFSET) :
+ thread_count;

This patch method refers to dmidecode, which is the same as it, like this:

                        if (data[0x25] != 0)
                                pr_attr("Thread Count", "%u",
                                        h->length >= 0x30 && data[0x25] == 0xFF ?

                                        WORD(data + 0x2E) : data[0x25]);


According SPEC, it uses THREAD_COUNT_2 only if dm->length >= 30 and
thread_count == 0xff, so the correct and simplest way is like this:

cpu_clock_freq = freq_temp * 1000000;

loongson_sysconf.cpuname = (void *)dmi_string_parse(dm, dmi_data[16]);
- loongson_sysconf.cores_per_package = *(dmi_data +
SMBIOS_THREAD_PACKAGE_OFFSET);
+ if (dm->length < 0x30)
+ thread_count = *(u8 *)(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
+ else if (thread_count == 0xff)
+ thread_count = *(u16 *)(dmi_data +
SMBIOS_THREAD_PACKAGE_2_OFFSET);
+ loongson_sysconf.cores_per_package = thread_count;

pr_info("CpuClock = %llu\n", cpu_clock_freq);
I feel there is a problem with this method.
I saw on the 3C6000/Q machine a few days ago that the dm->lenght is 0x30,
but only one byte is used, and the thread_count is 128, not 0xFF.
Is the initialization thread_count = *(u8 *)(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET) missing?
Wouldn't it be better to change it this way?

+       u8 thread_count;

        freq_temp = ((*(dmi_data + SMBIOS_FREQHIGH_OFFSET) << 8) +
                        ((*(dmi_data + SMBIOS_FREQLOW_OFFSET)) & SMBIOS_FREQLOW_MASK));
        cpu_clock_freq = freq_temp * 1000000;

        loongson_sysconf.cpuname = (void *)dmi_string_parse(dm, dmi_data[16]);
-       loongson_sysconf.cores_per_package = *(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
+       thread_count = *(u8 *)(dmi_data + SMBIOS_THREAD_PACKAGE_OFFSET);
+       if (thread_count == 0xFF && dm->length >= 0x30)
+               thread_count = *(u16 *)(dmi_data + SMBIOS_THREAD_PACKAGE_2_OFFSET);
+       loongson_sysconf.cores_per_package = thread_count;

        pr_info("CpuClock = %llu\n", cpu_clock_freq);
 }
Huacai

pr_info("CpuClock = %llu\n", cpu_clock_freq);
}
--
2.20.1