On 25/03/2021 17:15, 丁宁(富邻) wrote: > I have a question here. If crc32c instruction on some Zhaoxin CPUs > slower than generic C > implementation, then why not clear the X86_FEATURE_XMM4_2 flag? > The presence of crc32c and some other instructions supports are enumerated by CPUID.01:ECX[SSE4.2] = 1 that called X86_FEATURE_XMM4_2 in kernel, other instructions are ok except the crc32c instruction on these Zhaoxin CPUs. Sincerely TonyWWangoc

> Best Regards > Artie Ding（丁宁） >    ------------------原始邮件 ------------------ > 发件人:Tony W Wang-oc &lt;TonyWWang-oc(a)zhaoxin.com&gt; > 发送时间:Thu Mar 25 16:37:49 2021 > 收件人:丁宁(富邻) &lt;fulin.dn(a)alibaba-inc.com&gt;om>, LeoLiu-oc > &lt;LeoLiu-oc(a)zhaoxin.com&gt;om>, cloud-kernel > &lt;cloud-kernel(a)lists.openanolis.org&gt; > 抄送:TimGuo &lt;TimGuo(a)zhaoxin.com&gt;om>, CobeChen &lt;CobeChen(a)zhaoxin.com&gt; > 主题:Re: 回复：[ck] [patch v1] ck: crypto: x86/crc32c-intel - Don't > match some Zhaoxin CPUs > On 25/03/2021 15:53, 丁宁(富邻) wrote: >> ​If X86_FEATURE_XMM4_2 meens different hw implement to Zhaoxin CPUs, >> a new flag is needed to distinguish >> them. >> > The X86_FEATURE_XMM4_2 on Zhaoxin CPUs have the same meaning with > Intel. > crc32c instruction on some Zhaoxin CPUs slower than generic C > implementation, should avoid these Zhaoxin CPUs using the driver > crc32c-intel which will call crc32c instruction. > > Sincerely > TonyWWangoc > >> Artie >> >> >>    ------------------原始邮件 ------------------ >> 发件人:LeoLiu-oc &lt;LeoLiu-oc(a)zhaoxin.com&gt; >> 发送时间:Mon Mar 22 18:44:54 2021 >> 收件人: &lt;cloud-kernel(a)lists.openanolis.org&gt; >> 抄送: &lt;TonyWWang-oc(a)zhaoxin.com&gt;om>, &lt;TimGuo(a)zhaoxin.com&gt;om>, >> &lt;CobeChen(a)zhaoxin.com&gt; >> 主题:[ck] [patch v1] ck: crypto: x86/crc32c-intel - Don't match some >> Zhaoxin CPUs >> >> The driver crc32c-intel match CPUs supporting X86_FEATURE_XMM4_2. >> On platforms with Zhaoxin CPUs supporting this X86 feature, >> when crc32c-intel and crc32c-generic are both registered, >> system will use crc32c-intel because its .cra_priority is greater >> than crc32c-generic. >> >> When doing lmbench3 Create and Delete file test on partitions with >> ext4 enabling metadata checksum, found using crc32c-generic driver >> could get about 20% performance gain than using the driver >> crc32c-intel on some Zhaoxin CPUs. >> >> This case expect to use crc32c-generic driver for these Zhaoxin CPUs >> to get performance gain, so remove these Zhaoxin CPUs support from >> crc32c-intel. >> >> This patch was submitted to mainline kernel but not accepted by >> upstream >> maintainer whose reason is "Then create a BUG flag for it,". >> >> We think this is not a CPU bug for Zhaoxin CPUs. So should patch the >> crc32c driver for Zhaoxin CPUs but not report a BUG. >> >> https://lkml.org/lkml/2020/12/11/308 >> >> Signed-off-by: LeoLiu-oc &lt;LeoLiu-oc(a)zhaoxin.com&gt; >> --- >>     arch/x86/crypto/crc32c-intel_glue.c | 5 +++++ >>     1 file changed, 5 insertions(+) >> >> diff --git a/arch/x86/crypto/crc32c-intel_glue.c >> b/arch/x86/crypto/crc32c-intel_glue.c >> index 5773e1161072..d994bd49761a 100644 >> --- a/arch/x86/crypto/crc32c-intel_glue.c >> +++ b/arch/x86/crypto/crc32c-intel_glue.c >> @@ -242,8 +242,13 @@ MODULE_DEVICE_TABLE(x86cpu, crc32c_cpu_id); >> >>     static int __init crc32c_intel_mod_init(void) >>     { >> +struct cpuinfo_x86 *c = &boot_cpu_data; >>     if (!x86_match_cpu(crc32c_cpu_id)) >>     return -ENODEV; >> +if ((c->x86_vendor == X86_VENDOR_ZHAOXIN || c->x86_vendor == >> X86_VENDOR_CENTAUR) && >> +(c->x86 <= 7 && c->x86_model <= 59)) { >> +return -ENODEV; >> +} >>     #ifdef CONFIG_X86_64 >>     if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) { >>     alg.update = crc32c_pcl_intel_update;