On Thu, 2013-07-18 at 12:47 +0900, Tetsuo Handa wrote:
> Tim Chen wrote:
> > > > Your approach is quite complicated. I think something simpler like the
> > > > following will work:
> > >
> > > We cannot benefit from PCLMULQDQ. Is it acceptable for you?
> >
> >
> > The following code in crct10dif-pclmul_glue.c
> >
> > static const struct x86_cpu_id crct10dif_cpu_id[] = {
> > X86_FEATURE_MATCH(X86_FEATURE_PCLMULQDQ),
> > {}
> > };
> > MODULE_DEVICE_TABLE(x86cpu, crct10dif_cpu_id);
> >
> > will put the module in the device table and get the module
> > loaded, as long as the cpu support PCLMULQDQ. So we should be able
> > to benefit.
>
> Excuse me, how can crct10dif-pclmul.ko get loaded automatically?
> Did you test CONFIG_CRYPTO_CRCT10DIF_PCLMUL=m with below debug message?
The code:
static const struct x86_cpu_id crct10dif_cpu_id[] = {
X86_FEATURE_MATCH(X86_FEATURE_PCLMULQDQ),
{}
};
MODULE_DEVICE_TABLE(x86cpu, crct10dif_cpu_id);
causes the following line to be added to modules.alias file:
alias x86cpu:vendor:*:family:*:model:*:feature:*0081* crct10dif_pclmul
This should cause udev to load the crct10dif_pclml module when cpu
support the PCLMULQDQ (feature code 0081). I did my testing during
development on 3.10 and the module was indeed loaded.
However, I found that the following commit under 3.11-rc1 broke
the mechanism after some bisection.
commit ac212b6980d8d5eda705864fc5a8ecddc6d6eacc
Author: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
Date: Fri May 3 00:26:22 2013 +0200
ACPI / processor: Use common hotplug infrastructure
Split the ACPI processor driver into two parts, one that is
non-modular, resides in the ACPI core and handles the enumeration
and hotplug of processors and one that implements the rest of the
existing processor driver functionality.
Rafael, can you check and see if this can be fixed so those optimized
crypto modules for Intel cpu that support them can be loaded?
Herbert, we had a discussion earlier on a separate issue regarding the
module load order. We wanted to load all supported crypto t10dif modules
before the crc-t10dif module. You mentioned that the MODULE_ALIAS also
need some fixing and wonder if those changes have been merged into
3.11-rc1? See https://lkml.org/lkml/2013/5/21/216 . Theoretically, that
fix should also get all the crypto modules support t10dif be loaded.
Thanks.
Tim
>
> diff --git a/arch/x86/crypto/crct10dif-pclmul_glue.c b/arch/x86/crypto/crct10dif-pclmul_glue.c
> index 7845d7f..a8a95aa 100644
> --- a/arch/x86/crypto/crct10dif-pclmul_glue.c
> +++ b/arch/x86/crypto/crct10dif-pclmul_glue.c
> @@ -129,9 +129,10 @@ MODULE_DEVICE_TABLE(x86cpu, crct10dif_cpu_id);
>
> static int __init crct10dif_intel_mod_init(void)
> {
> + printk(KERN_WARNING "********** Checking for X86_FEATURE_PCLMULQDQ\n");
> if (!x86_match_cpu(crct10dif_cpu_id))
> return -ENODEV;
> -
> + printk(KERN_WARNING "********** Registering crct10dif-pclmul\n");
> return crypto_register_shash(&alg);
> }
>
> As far as I tested, crct10dif-pclmul.ko will not be loaded unless manually
> adding "modprobe crct10dif-pclmul" to initramfs's /init or choosing
> CONFIG_CRYPTO_CRCT10DIF_PCLMUL=y.
>
> > So as long as the crct10dif.ko and crct10dif-pclmul.ko are loaded,
> > the pclmulqdq t10dif will have a higher priority and get allocated
> > and used.
>
> What I'm talking are
>
> (1) Since mkinitramfs is unable to know that crct10dif-pclmul.ko has higher
> priority than crct10dif.ko , mkinitramfs will not include
> "modprobe crct10dif-pclmul" line in the generated initramfs.
>
> (2) In order to get benefit from PCLMULQDQ, users have to manually make sure
> that "modprobe crct10dif-pclmul" is called before crc-t10dif.ko (which is
> loaded before sd_mod.ko is loaded) is loaded by their initramfs's /init
> script.
>
> (3) The cause of (1) is that crct10dif-pclmul.ko will not be loaded
> automatically unless choosing CONFIG_CRYPTO_CRCT10DIF_PCLMUL=y.
>
> (4) The cause of (3) is that modules.dep does not describe that users will
> benefit by loading crct10dif-pclmul.ko before loading crc-t10dif.ko .
>
> (5) Currently crct10dif-pclmul.ko cannot be loaded if PCLMULQDQ is not
> supported. This leads to boot failure (since sd_mod.ko cannot be loaded)
> if modules.dep says that "crct10dif-pclmul.ko is required by
> crc-t10dif.ko".
>
> (6) To solve (4) and (5), modules.dep should say "crct10dif-pclmul.ko is
> preferred for crc-t10dif.ko but is not required by crc-t10dif.ko".
> But there is no such mechanism. Thus, currently available choice is
> "allow loading crct10dif-pclmul.ko even if PCLMULQDQ is not supported"
> or "ignore errors by built-in the crct10dif-pclmul.ko module".
>
> My patch (b) seems to be complicated but is required in order to solve (4)
> without asking users to manually add "modprobe crct10dif-pclmul" into their
> initramfs. If we choose patch (b) rather than patch (a), we need to solve (5).
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
