Re: [patch 05/15] x86/tlb: Move __flush_tlb() out of line

[Tom Lendacky <thomas.lendacky@amd.com>]

On 4/20/20 9:38 AM, Jürgen Groß wrote:
> On 20.04.20 16:26, Tom Lendacky wrote:
>> On 4/20/20 9:03 AM, Jürgen Groß wrote:
>>> On 20.04.20 15:48, Tom Lendacky wrote:
>>>> On 4/19/20 3:31 PM, Thomas Gleixner wrote:
>>>>> cpu_tlbstate is exported because various TLB related functions need 
>>>>> access
>>>>> to it, but cpu_tlbstate is sensitive information which should only be
>>>>> accessed by well contained kernel functions and not be directly 
>>>>> exposed to
>>>>> modules.
>>>>>
>>>>> The various TLB flush functions need access to cpu_tlbstate. As a first
>>>>> step move __flush_tlb() out of line and hide the native function. The
>>>>> latter can be static when CONFIG_PARAVIRT is disabled.
>>>>>
>>>>> Consolidate the name space while at it and remove the pointless extra
>>>>> wrapper in the paravirt code.
>>>>>
>>>>> No functional change.
>>>>>
>>>>> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
>>>>> Cc: Thomas Lendacky <Thomas.Lendacky@amd.com>
>>>>> Cc: Juergen Gross <jgross@suse.com>
>>>>> ---
>>>>>   arch/x86/include/asm/paravirt.h    |    4 +++-
>>>>>   arch/x86/include/asm/tlbflush.h    |   29 
>>>>> +++++------------------------
>>>>>   arch/x86/kernel/cpu/mtrr/generic.c |    4 ++--
>>>>>   arch/x86/kernel/paravirt.c         |    7 +------
>>>>>   arch/x86/mm/mem_encrypt.c          |    2 +-
>>>>>   arch/x86/mm/tlb.c                  |   33 
>>>>> ++++++++++++++++++++++++++++++++-
>>>>>   arch/x86/platform/uv/tlb_uv.c      |    2 +-
>>>>>   7 files changed, 45 insertions(+), 36 deletions(-)
>>>>>
>>>>> --- a/arch/x86/include/asm/paravirt.h
>>>>> +++ b/arch/x86/include/asm/paravirt.h
>>>>> @@ -47,7 +47,9 @@ static inline void slow_down_io(void)
>>>>>   #endif
>>>>>   }
>>>>> -static inline void __flush_tlb(void)
>>>>> +void native_flush_tlb_local(void);
>>>>> +
>>>>> +static inline void __flush_tlb_local(void)
>>>>>   {
>>>>>       PVOP_VCALL0(mmu.flush_tlb_user);
>>>>>   }
>>>>> --- a/arch/x86/include/asm/tlbflush.h
>>>>> +++ b/arch/x86/include/asm/tlbflush.h
>>>>> @@ -140,12 +140,13 @@ static inline unsigned long build_cr3_no
>>>>>       return __sme_pa(pgd) | kern_pcid(asid) | CR3_NOFLUSH;
>>>>>   }
>>>>> +void flush_tlb_local(void);
>>>>> +
>>>>>   #ifdef CONFIG_PARAVIRT
>>>>>   #include <asm/paravirt.h>
>>>>>   #else
>>>>> -#define __flush_tlb() __native_flush_tlb()
>>>>> -#define __flush_tlb_global() __native_flush_tlb_global()
>>>>> -#define __flush_tlb_one_user(addr) __native_flush_tlb_one_user(addr)
>>>>> +#define __flush_tlb_global()        __native_flush_tlb_global()
>>>>> +#define __flush_tlb_one_user(addr) __native_flush_tlb_one_user(addr)
>>>>>   #endif
>>>>>   struct tlb_context {
>>>>> @@ -371,24 +372,6 @@ static inline void invalidate_user_asid(
>>>>>   }
>>>>>   /*
>>>>> - * flush the entire current user mapping
>>>>> - */
>>>>> -static inline void __native_flush_tlb(void)
>>>>> -{
>>>>> -    /*
>>>>> -     * Preemption or interrupts must be disabled to protect the access
>>>>> -     * to the per CPU variable and to prevent being preempted between
>>>>> -     * read_cr3() and write_cr3().
>>>>> -     */
>>>>> -    WARN_ON_ONCE(preemptible());
>>>>> -
>>>>> -    invalidate_user_asid(this_cpu_read(cpu_tlbstate.loaded_mm_asid));
>>>>> -
>>>>> -    /* If current->mm == NULL then the read_cr3() "borrows" an mm */
>>>>> -    native_write_cr3(__native_read_cr3());
>>>>> -}
>>>>> -
>>>>> -/*
>>>>>    * flush everything
>>>>>    */
>>>>>   static inline void __native_flush_tlb_global(void)
>>>>> @@ -461,7 +444,7 @@ static inline void __flush_tlb_all(void)
>>>>>           /*
>>>>>            * !PGE -> !PCID (setup_pcid()), thus every flush is total.
>>>>>            */
>>>>> -        __flush_tlb();
>>>>> +        flush_tlb_local();
>>>>>       }
>>>>>   }
>>>>> @@ -537,8 +520,6 @@ struct flush_tlb_info {
>>>>>       bool            freed_tables;
>>>>>   };
>>>>> -#define local_flush_tlb() __flush_tlb()
>>>>> -
>>>>>   #define flush_tlb_mm(mm)                        \
>>>>>           flush_tlb_mm_range(mm, 0UL, TLB_FLUSH_ALL, 0UL, true)
>>>>> --- a/arch/x86/kernel/cpu/mtrr/generic.c
>>>>> +++ b/arch/x86/kernel/cpu/mtrr/generic.c
>>>>> @@ -761,7 +761,7 @@ static void prepare_set(void) __acquires
>>>>>       /* Flush all TLBs via a mov %cr3, %reg; mov %reg, %cr3 */
>>>>>       count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
>>>>> -    __flush_tlb();
>>>>> +    flush_tlb_local();
>>>>>       /* Save MTRR state */
>>>>>       rdmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
>>>>> @@ -778,7 +778,7 @@ static void post_set(void) __releases(se
>>>>>   {
>>>>>       /* Flush TLBs (no need to flush caches - they are disabled) */
>>>>>       count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
>>>>> -    __flush_tlb();
>>>>> +    flush_tlb_local();
>>>>>       /* Intel (P6) standard MTRRs */
>>>>>       mtrr_wrmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
>>>>> --- a/arch/x86/kernel/paravirt.c
>>>>> +++ b/arch/x86/kernel/paravirt.c
>>>>> @@ -160,11 +160,6 @@ unsigned paravirt_patch_insns(void *insn
>>>>>       return insn_len;
>>>>>   }
>>>>> -static void native_flush_tlb(void)
>>>>> -{
>>>>> -    __native_flush_tlb();
>>>>> -}
>>>>> -
>>>>>   /*
>>>>>    * Global pages have to be flushed a bit differently. Not a real
>>>>>    * performance problem because this does not happen often.
>>>>> @@ -359,7 +354,7 @@ struct paravirt_patch_template pv_ops =
>>>>>   #endif /* CONFIG_PARAVIRT_XXL */
>>>>>       /* Mmu ops. */
>>>>> -    .mmu.flush_tlb_user    = native_flush_tlb,
>>>>> +    .mmu.flush_tlb_user    = native_flush_tlb_local,
>>>>>       .mmu.flush_tlb_kernel    = native_flush_tlb_global,
>>>>>       .mmu.flush_tlb_one_user    = native_flush_tlb_one_user,
>>>>>       .mmu.flush_tlb_others    = native_flush_tlb_others,
>>>>> --- a/arch/x86/mm/mem_encrypt.c
>>>>> +++ b/arch/x86/mm/mem_encrypt.c
>>>>> @@ -134,7 +134,7 @@ static void __init __sme_early_map_unmap
>>>>>           size = (size <= PMD_SIZE) ? 0 : size - PMD_SIZE;
>>>>>       } while (size);
>>>>> -    __native_flush_tlb();
>>>>> +    flush_tlb_local();
>>>>
>>>> This invoked __native_flush_tlb() because of how early it is called 
>>>> and the paravirt ops support isn't set up yet, resulting in a crash if 
>>>> not invoking the native version directly. So this needs a "native" 
>>>> version of the tlb flush to invoke.
>>>
>>> I don't think this is still true. With my rework of pvops to have all
>>> functions in one struct which is initialized statically initially
>>> everything should work from the time the kernel is mapped.
>>>
>>> In case it doesn't there is something very wrong IMO.
>>
>> The memory encryption support was implemented in 4.14, so it's quite 
>> possible that this isn't an issue now. I'll test out the patch and 
>> verify it. What release did your pvops rework land in?
> 
> 4.20.

Ok, my (limited) testing appears good at this point. I see a couple of 
calls to ftrace (__fentry__) that just return before invoking the native 
version of the TLB flushing function.

I don't remember what the original problem was that required the call to 
__native_flush_tlb(), maybe some overall kernel instrumentation that once 
removed via the Makefile took care of it. But it appears ok now, so 
disregard my original comment.

Thanks,
Tom

> 
> 
> Juergen
>