Re: [PATCH v15 09/10] arm64: Call kmalloc() to allocate DTB buffer

From: Lakshmi Ramasubramanian <nramas@linux.microsoft.com>
To: Thiago Jung Bauermann <bauerman@linux.ibm.com>,
	Will Deacon <will@kernel.org>
Cc: zohar@linux.ibm.com, robh@kernel.org, takahiro.akashi@linaro.org,
	gregkh@linuxfoundation.org, catalin.marinas@arm.com,
	mpe@ellerman.id.au, james.morse@arm.com, sashal@kernel.org,
	benh@kernel.crashing.org, paulus@samba.org,
	frowand.list@gmail.com, vincenzo.frascino@arm.com,
	mark.rutland@arm.com, dmitry.kasatkin@gmail.com,
	jmorris@namei.org, serge@hallyn.com, pasha.tatashin@soleen.com,
	allison@lohutok.net, masahiroy@kernel.org, bhsharma@redhat.com,
	mbrugger@suse.com, hsinyi@chromium.org, tao.li@vivo.com,
	christophe.leroy@c-s.fr, prsriva@linux.microsoft.com,
	balajib@linux.microsoft.com, linux-integrity@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org, devicetree@vger.kernel.org,
	linuxppc-dev@lists.ozlabs.org
Subject: Re: [PATCH v15 09/10] arm64: Call kmalloc() to allocate DTB buffer
Date: Wed, 27 Jan 2021 20:00:34 -0800	[thread overview]
Message-ID: <58d3ffbf-4d80-c893-34d6-366ebfac55bd@linux.microsoft.com> (raw)
In-Reply-To: <871re5soof.fsf@manicouagan.localdomain>

On 1/27/21 7:52 PM, Thiago Jung Bauermann wrote:
> 
> Will Deacon <will@kernel.org> writes:
> 
>> On Wed, Jan 27, 2021 at 09:59:38AM -0800, Lakshmi Ramasubramanian wrote:
>>> On 1/27/21 8:52 AM, Will Deacon wrote:
>>>
>>> Hi Will,
>>>
>>>> On Fri, Jan 15, 2021 at 09:30:16AM -0800, Lakshmi Ramasubramanian wrote:
>>>>> create_dtb() function allocates kernel virtual memory for
>>>>> the device tree blob (DTB).  This is not consistent with other
>>>>> architectures, such as powerpc, which calls kmalloc() for allocating
>>>>> memory for the DTB.
>>>>>
>>>>> Call kmalloc() to allocate memory for the DTB, and kfree() to free
>>>>> the allocated memory.
>>>>>
>>>>> Co-developed-by: Prakhar Srivastava <prsriva@linux.microsoft.com>
>>>>> Signed-off-by: Prakhar Srivastava <prsriva@linux.microsoft.com>
>>>>> Signed-off-by: Lakshmi Ramasubramanian <nramas@linux.microsoft.com>
>>>>> ---
>>>>>    arch/arm64/kernel/machine_kexec_file.c | 12 +++++++-----
>>>>>    1 file changed, 7 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/arch/arm64/kernel/machine_kexec_file.c b/arch/arm64/kernel/machine_kexec_file.c
>>>>> index 7de9c47dee7c..51c40143d6fa 100644
>>>>> --- a/arch/arm64/kernel/machine_kexec_file.c
>>>>> +++ b/arch/arm64/kernel/machine_kexec_file.c
>>>>> @@ -29,7 +29,7 @@ const struct kexec_file_ops * const kexec_file_loaders[] = {
>>>>>    int arch_kimage_file_post_load_cleanup(struct kimage *image)
>>>>>    {
>>>>> -	vfree(image->arch.dtb);
>>>>> +	kfree(image->arch.dtb);
>>>>>    	image->arch.dtb = NULL;
>>>>>    	vfree(image->arch.elf_headers);
>>>>> @@ -59,19 +59,21 @@ static int create_dtb(struct kimage *image,
>>>>>    			+ cmdline_len + DTB_EXTRA_SPACE;
>>>>>    	for (;;) {
>>>>> -		buf = vmalloc(buf_size);
>>>>> +		buf = kmalloc(buf_size, GFP_KERNEL);
>>>>
>>>> Is there a functional need for this patch? I build the 'dtbs' target just
>>>> now and sdm845-db845c.dtb is approaching 100K, which feels quite large
>>>> for kmalloc().
>>>
>>> Changing the allocation from vmalloc() to kmalloc() would help us further
>>> consolidate the DTB setup code for powerpc and arm64.
>>
>> Ok, but at the risk of allocation failure. Can powerpc use vmalloc()
>> instead?
> 
> I believe this patch stems from this suggestion by Rob Herring:
> 
>> This could be taken a step further and do the allocation of the new
>> FDT. The difference is arm64 uses vmalloc and powerpc uses kmalloc. The
>> arm64 version also retries with a bigger allocation. That seems
>> unnecessary.
> 
> in https://lore.kernel.org/linux-integrity/20201211221006.1052453-3-robh@kernel.org/
> 
> The problem is that this patch implements only part of the suggestion,
> which isn't useful in itself. So the patch series should either drop
> this patch or consolidate the FDT allocation between the arches.
> 
> I just tested on powernv and pseries platforms and powerpc can use
> vmalloc for the FDT buffer.
> 

Thanks for verifying on powerpc platform Thiago.

I'll update the patch to do the following:

=> Use vmalloc for FDT buffer allocation on powerpc
=> Keep vmalloc for arm64, but remove the retry on allocation.
=> Also, there was a memory leak of FDT buffer in the error code path on 
arm64, which I'll fix as well.

Did I miss anything?

thanks,
  -lakshmi