Re: [PATCH v6] mm,kfence: decouple kfence from page granularity mapping judgement

[Zhenhua Huang <quic_zhenhuah@quicinc.com>]

On 2023/3/13 22:42, Kefeng Wang wrote:
> 
> 
> On 2023/3/13 21:00, Marco Elver wrote:
>> On Mon, 13 Mar 2023 at 12:57, Zhenhua Huang 
>> <quic_zhenhuah@quicinc.com> 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/mm/mmu.c      | 42 
>>> ++++++++++++++++++++++++++++++++++++++++++
>>>   arch/arm64/mm/pageattr.c |  8 ++++++--
>>>   include/linux/kfence.h   | 10 ++++++++++
>>>   mm/kfence/core.c         |  9 +++++++++
>>>   4 files changed, 67 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>>> index 6f9d889..ca5c932 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>
>>> @@ -525,6 +526,31 @@ 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;
>>> +
>>> +       kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
>>> +       if (!kfence_pool)
>>> +               pr_err("failed to allocate kfence pool\n");
>>> +
>>> +       return kfence_pool;
>>> +}
>>> +
>>> +#else
>>> +
>>> +static phys_addr_t arm64_kfence_alloc_pool(void)
>>> +{
>>> +       return 0;
>>> +}
>>> +
>>> +#endif
>>> +
>>>   static void __init map_mem(pgd_t *pgdp)
>>>   {
>>>          static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
>>> @@ -532,6 +558,7 @@ static void __init map_mem(pgd_t *pgdp)
>>>          phys_addr_t kernel_end = __pa_symbol(__init_begin);
>>>          phys_addr_t start, end;
>>>          int flags = NO_EXEC_MAPPINGS;
>>> +       phys_addr_t kfence_pool;
>>>          u64 i;
>>>
>>>          /*
>>> @@ -564,6 +591,10 @@ static void __init map_mem(pgd_t *pgdp)
>>>          }
>>>   #endif
>>>
>>> +       kfence_pool = arm64_kfence_alloc_pool();
>>> +       if (kfence_pool)
>>> +               memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE);
>>> +
>>>          /* map all the memory banks */
>>>          for_each_mem_range(i, &start, &end) {
>>>                  if (start >= end)
>>> @@ -608,6 +639,17 @@ static void __init map_mem(pgd_t *pgdp)
>>>                  }
>>>          }
>>>   #endif
>>> +
>>> +       /* Kfence pool needs page-level mapping */
>>> +       if (kfence_pool) {
>>> +               __map_memblock(pgdp, kfence_pool,
>>> +                       kfence_pool + KFENCE_POOL_SIZE,
>>> +                       pgprot_tagged(PAGE_KERNEL),
>>> +                       NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
>>> +               memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE);
>>> +               /* kfence_pool really mapped now */
>>> +               kfence_set_pool(kfence_pool);
>>> +       }
>>>   }
>>>
>>>   void mark_rodata_ro(void)
>>> diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
>>> index 79dd201..25e4a983 100644
>>> --- a/arch/arm64/mm/pageattr.c
>>> +++ b/arch/arm64/mm/pageattr.c
>>> @@ -7,6 +7,7 @@
>>>   #include <linux/module.h>
>>>   #include <linux/sched.h>
>>>   #include <linux/vmalloc.h>
>>> +#include <linux/kfence.h>
>>>
>>>   #include <asm/cacheflush.h>
>>>   #include <asm/set_memory.h>
>>> @@ -22,12 +23,15 @@ bool rodata_full __ro_after_init = 
>>> IS_ENABLED(CONFIG_RODATA_FULL_DEFAULT_ENABLED
>>>   bool can_set_direct_map(void)
>>>   {
>>>          /*
>>> -        * rodata_full, DEBUG_PAGEALLOC and KFENCE require linear map 
>>> to be
>>> +        * rodata_full and DEBUG_PAGEALLOC require linear map to be
>>>           * mapped at page granularity, so that it is possible to
>>>           * protect/unprotect single pages.
>>> +        *
>>> +        * Kfence pool requires page granularity mapping also if we 
>>> init it
>>> +        * late.
>>>           */
>>>          return (rodata_enabled && rodata_full) || 
>>> debug_pagealloc_enabled() ||
>>> -               IS_ENABLED(CONFIG_KFENCE);
>>> +           (IS_ENABLED(CONFIG_KFENCE) && !kfence_sample_interval);
>>
>> If you're struggling with kfence_sample_interval not existing if
>> !CONFIG_KFENCE, this is one of the occasions where it'd be perfectly
>> fine to write:
>>
>> bool can_set_direct_map(void) {
>> #ifdef CONFIG_KFENCE
>>      /* ... your comment here ...*/
>>      if (!kfence_sample_interval)
>>          return true;
>> }
>> #endif
>>       return .........
>> }
>>
>>>   }
>>>
> The can_set_direct_map() could be called anytime, eg, memory add,
> vmalloc, and this will make different state of can_set_direct_map()
> if kfence is re-enabled, I think that we need a new value to check 
> whether or not the early kfence_pool is initialized.

Many thanks, Kefeng and Marco for your careful review. Agree, 
kfence_sample_interval can be modified in a few ways and we can't use it 
in can_set_direct_map().

To be honest, previously I wanted to allocate kfence pool early always 
but it seems breaks the flexibility that b33f778bba5e ("kfence: alloc 
kfence_pool after system startup") introduced.

Now I prefer to introduce one global variable early_kfence_pool to 
indicate if kfence_pool is initialized early, then can_set_direct_map() 
should be easy and clear to handle: just add "(IS_ENABLED(CONFIG_KFENCE) 
&& !early_kfence_pool)" for the case of possibility we may init kfence 
pool later. The naming of early_kfence_pool also can well expressed what 
we're doing :)

How about your idea? I will update a new patchset.

> 
>