Re: [PATCH v8] mm,kfence: decouple kfence from page granularity mapping judgement

[Zhenhua Huang <quic_zhenhuah@quicinc.com>]

Thanks Pavan.

On 2023/3/14 19:14, Pavan Kondeti wrote:
> On Tue, Mar 14, 2023 at 06:08:07PM +0800, Zhenhua Huang wrote:
>>
>>
>> On 2023/3/14 16:36, Pavan Kondeti wrote:
>>> On Tue, Mar 14, 2023 at 03:05:02PM +0800, Zhenhua Huang wrote:
>>>> Kfence only needs its pool to be mapped as page granularity, if it is
>>>> inited early. Previous judgement was a bit over protected. From [1], Mark
>>>> suggested to "just map the KFENCE region a page granularity". So I
>>>> decouple it from judgement and do page granularity mapping for kfence
>>>> pool only. Need to be noticed that late init of kfence pool still requires
>>>> page granularity mapping.
>>>>
>>>> Page granularity mapping in theory cost more(2M per 1GB) memory on arm64
>>>> platform. Like what I've tested on QEMU(emulated 1GB RAM) with
>>>> gki_defconfig, also turning off rodata protection:
>>>> Before:
>>>> [root@liebao ]# cat /proc/meminfo
>>>> MemTotal:         999484 kB
>>>> After:
>>>> [root@liebao ]# cat /proc/meminfo
>>>> MemTotal:        1001480 kB
>>>>
>>>> To implement this, also relocate the kfence pool allocation before the
>>>> linear mapping setting up, arm64_kfence_alloc_pool is to allocate phys
>>>> addr, __kfence_pool is to be set after linear mapping set up.
>>>>
>>>> LINK: [1] https://lore.kernel.org/linux-arm-kernel/Y+IsdrvDNILA59UN@FVFF77S0Q05N/
>>>> Suggested-by: Mark Rutland <mark.rutland@arm.com>
>>>> Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com>
>>>> ---
>>>>    arch/arm64/include/asm/kfence.h |  2 ++
>>>>    arch/arm64/mm/mmu.c             | 44 +++++++++++++++++++++++++++++++++++++++++
>>>>    arch/arm64/mm/pageattr.c        |  9 +++++++--
>>>>    include/linux/kfence.h          |  8 ++++++++
>>>>    mm/kfence/core.c                |  9 +++++++++
>>>>    5 files changed, 70 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/include/asm/kfence.h b/arch/arm64/include/asm/kfence.h
>>>> index aa855c6..f1f9ca2d 100644
>>>> --- a/arch/arm64/include/asm/kfence.h
>>>> +++ b/arch/arm64/include/asm/kfence.h
>>>> @@ -10,6 +10,8 @@
>>>>    #include <asm/set_memory.h>
>>>> +extern phys_addr_t early_kfence_pool;
>>>> +
>>>>    static inline bool arch_kfence_init_pool(void) { return true; }
>>>>    static inline bool kfence_protect_page(unsigned long addr, bool protect)
>>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>>>> index 6f9d889..7fbf2ed 100644
>>>> --- a/arch/arm64/mm/mmu.c
>>>> +++ b/arch/arm64/mm/mmu.c
>>>> @@ -24,6 +24,7 @@
>>>>    #include <linux/mm.h>
>>>>    #include <linux/vmalloc.h>
>>>>    #include <linux/set_memory.h>
>>>> +#include <linux/kfence.h>
>>>>    #include <asm/barrier.h>
>>>>    #include <asm/cputype.h>
>>>> @@ -38,6 +39,7 @@
>>>>    #include <asm/ptdump.h>
>>>>    #include <asm/tlbflush.h>
>>>>    #include <asm/pgalloc.h>
>>>> +#include <asm/kfence.h>
>>>>    #define NO_BLOCK_MAPPINGS	BIT(0)
>>>>    #define NO_CONT_MAPPINGS	BIT(1)
>>>> @@ -525,6 +527,33 @@ static int __init enable_crash_mem_map(char *arg)
>>>>    }
>>>>    early_param("crashkernel", enable_crash_mem_map);
>>>> +#ifdef CONFIG_KFENCE
>>>> +
>>>> +static phys_addr_t arm64_kfence_alloc_pool(void)
>>>> +{
>>>> +	phys_addr_t kfence_pool;
>>>> +
>>>> +	if (!kfence_sample_interval)
>>>> +		return 0;
>>>> +
>>>
>>> Are you sure that kernel commandline param are processed this early?
>>> AFAICS, start_kernel()->parse_args() process the kernel arguments. We
>>> are here before that. without your patch, mm_init() which takes care of
>>> allocating kfence memory is called after parse_args().
>>>
>>> Can you check your patch with kfence.sample_interval=0 appended to
>>> kernel commandline?
>>>
>>
>> Thanks Pavan. I have tried and you're correct. Previously I thought it's
>> parsed by the way:
>> setup_arch()->parse_early_param(earlier)->parse_early_options->
>> do_early_param
>> Unfortunately seems not take effect.
>>
>> Then the only way left is we always allocate the kfence pool early? as we
>> can't get sample_invertal at this early stage.
>>
> 
> That would mean, we would allocate the kfence pool memory even when it
> is disabled from commandline. That does not sound good to me.
> 
> Is it possible to free this early allocated memory later in
> mm_init()->kfence_alloc_pool()? if that is not possible, can we think of
> adding early param for kfence?

If we freed that buffer, there may be no chance to get that page 
granularity mapped buffer again.. as all these allocation/free are 
through normal buddy allocator.

At this stage, seems only additional early param can work.. Marco 
previously wanted to reuse sample_interval but seems not doable now.

Hi Marco,

Sorry, Can we thought of the solution again? like
ARM64:
1. intercepts early boot arg and gives early alloc memory to KFENCE
2. KFENCE to disable dynamic switch
3. disable page gran and save memory overhead
The purpose is in the case of w/o boot arg, it's just same as now.. arch 
specific kfence buffer will not allocate. And w/ boot arg, we can get 
expected saving.

Thanks,
Zhenhua

> 
>>>> +	kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
>>>> +	if (!kfence_pool)
>>>> +		pr_err("failed to allocate kfence pool\n");
>>>> +
>>> For whatever reason, if this allocation fails, what should be done? We
>>> end up not calling kfence_set_pool(). kfence_alloc_pool() is going to
>>> attempt allocation again but we did not setup page granularity. That
>>> means, we are enabling KFENCE without meeting pre-conditions. Can you
>>> check this?
>>
>> In this scenario, early_kfence_pool should be false(0) and we will end up
>> using page granularity mapping? should be fine IMO.
>>
> 
> Right, I missed that hunk in can_set_direct_map().
> 
> Thanks,
> Pavan