Message-Id: <7F3235AF-5300-49B6-A9A4-C8BA8EF1D87B@gmail.com>
Date:   Tue, 28 Aug 2018 10:33:02 -0700
From:   Nadav Amit <nadav.amit@...il.com>
To:     Masami Hiramatsu <mhiramat@...nel.org>,
        Andy Lutomirski <luto@...nel.org>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Kees Cook <keescook@...omium.org>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Jiri Kosina <jkosina@...e.cz>,
        Will Deacon <will.deacon@....com>,
        Benjamin Herrenschmidt <benh@....ibm.com>,
        Nick Piggin <npiggin@...il.com>,
        the arch/x86 maintainers <x86@...nel.org>,
        Borislav Petkov <bp@...en8.de>,
        Rik van Riel <riel@...riel.com>,
        Jann Horn <jannh@...gle.com>,
        Adin Scannell <ascannell@...gle.com>,
        Dave Hansen <dave.hansen@...el.com>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-mm <linux-mm@...ck.org>,
        David Miller <davem@...emloft.net>,
        Martin Schwidefsky <schwidefsky@...ibm.com>,
        Michael Ellerman <mpe@...erman.id.au>
Subject: Re: TLB flushes on fixmap changes

at 1:49 AM, Masami Hiramatsu <mhiramat@...nel.org> wrote:

> On Mon, 27 Aug 2018 16:01:32 -0700
> Andy Lutomirski <luto@...nel.org> wrote:
> 
>> On Mon, Aug 27, 2018 at 3:54 PM, Nadav Amit <nadav.amit@...il.com> wrote:
>>> at 3:32 PM, Andy Lutomirski <luto@...nel.org> wrote:
>>> 
>>>> On Mon, Aug 27, 2018 at 2:55 PM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>> at 1:16 PM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>> 
>>>>>> at 12:58 PM, Andy Lutomirski <luto@...nel.org> wrote:
>>>>>> 
>>>>>>> On Mon, Aug 27, 2018 at 12:43 PM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>>>>> at 12:10 PM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>>>>> 
>>>>>>>>> at 11:58 AM, Andy Lutomirski <luto@...nel.org> wrote:
>>>>>>>>> 
>>>>>>>>>> On Mon, Aug 27, 2018 at 11:54 AM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>>>>>>>>> On Mon, Aug 27, 2018 at 10:34 AM, Nadav Amit <nadav.amit@...il.com> wrote:
>>>>>>>>>>>> What do you all think?
>>>>>>>>>>> 
>>>>>>>>>>> I agree in general. But I think that current->mm would need to be loaded, as
>>>>>>>>>>> otherwise I am afraid it would break switch_mm_irqs_off().
>>>>>>>>>> 
>>>>>>>>>> What breaks?
>>>>>>>>> 
>>>>>>>>> Actually nothing. I just saw the IBPB stuff regarding tsk, but it should not
>>>>>>>>> matter.
>>>>>>>> 
>>>>>>>> So here is what I got. It certainly needs some cleanup, but it boots.
>>>>>>>> 
>>>>>>>> Let me know how crappy you find it...
>>>>>>>> 
>>>>>>>> 
>>>>>>>> diff --git a/arch/x86/include/asm/mmu_context.h b/arch/x86/include/asm/mmu_context.h
>>>>>>>> index bbc796eb0a3b..336779650a41 100644
>>>>>>>> --- a/arch/x86/include/asm/mmu_context.h
>>>>>>>> +++ b/arch/x86/include/asm/mmu_context.h
>>>>>>>> @@ -343,4 +343,24 @@ static inline unsigned long __get_current_cr3_fast(void)
>>>>>>>>     return cr3;
>>>>>>>> }
>>>>>>>> 
>>>>>>>> +typedef struct {
>>>>>>>> +       struct mm_struct *prev;
>>>>>>>> +} temporary_mm_state_t;
>>>>>>>> +
>>>>>>>> +static inline temporary_mm_state_t use_temporary_mm(struct mm_struct *mm)
>>>>>>>> +{
>>>>>>>> +       temporary_mm_state_t state;
>>>>>>>> +
>>>>>>>> +       lockdep_assert_irqs_disabled();
>>>>>>>> +       state.prev = this_cpu_read(cpu_tlbstate.loaded_mm);
>>>>>>>> +       switch_mm_irqs_off(NULL, mm, current);
>>>>>>>> +       return state;
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static inline void unuse_temporary_mm(temporary_mm_state_t prev)
>>>>>>>> +{
>>>>>>>> +       lockdep_assert_irqs_disabled();
>>>>>>>> +       switch_mm_irqs_off(NULL, prev.prev, current);
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> #endif /* _ASM_X86_MMU_CONTEXT_H */
>>>>>>>> diff --git a/arch/x86/include/asm/pgtable.h b/arch/x86/include/asm/pgtable.h
>>>>>>>> index 5715647fc4fe..ef62af9a0ef7 100644
>>>>>>>> --- a/arch/x86/include/asm/pgtable.h
>>>>>>>> +++ b/arch/x86/include/asm/pgtable.h
>>>>>>>> @@ -976,6 +976,10 @@ static inline void __meminit init_trampoline_default(void)
>>>>>>>>     /* Default trampoline pgd value */
>>>>>>>>     trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
>>>>>>>> }
>>>>>>>> +
>>>>>>>> +void __init patching_mm_init(void);
>>>>>>>> +#define patching_mm_init patching_mm_init
>>>>>>>> +
>>>>>>>> # ifdef CONFIG_RANDOMIZE_MEMORY
>>>>>>>> void __meminit init_trampoline(void);
>>>>>>>> # else
>>>>>>>> diff --git a/arch/x86/include/asm/pgtable_64_types.h b/arch/x86/include/asm/pgtable_64_types.h
>>>>>>>> index 054765ab2da2..9f44262abde0 100644
>>>>>>>> --- a/arch/x86/include/asm/pgtable_64_types.h
>>>>>>>> +++ b/arch/x86/include/asm/pgtable_64_types.h
>>>>>>>> @@ -116,6 +116,9 @@ extern unsigned int ptrs_per_p4d;
>>>>>>>> #define LDT_PGD_ENTRY          (pgtable_l5_enabled() ? LDT_PGD_ENTRY_L5 : LDT_PGD_ENTRY_L4)
>>>>>>>> #define LDT_BASE_ADDR          (LDT_PGD_ENTRY << PGDIR_SHIFT)
>>>>>>>> 
>>>>>>>> +#define TEXT_POKE_PGD_ENTRY    -5UL
>>>>>>>> +#define TEXT_POKE_ADDR         (TEXT_POKE_PGD_ENTRY << PGDIR_SHIFT)
>>>>>>>> +
>>>>>>>> #define __VMALLOC_BASE_L4      0xffffc90000000000UL
>>>>>>>> #define __VMALLOC_BASE_L5      0xffa0000000000000UL
>>>>>>>> 
>>>>>>>> diff --git a/arch/x86/include/asm/pgtable_types.h b/arch/x86/include/asm/pgtable_types.h
>>>>>>>> index 99fff853c944..840c72ec8c4f 100644
>>>>>>>> --- a/arch/x86/include/asm/pgtable_types.h
>>>>>>>> +++ b/arch/x86/include/asm/pgtable_types.h
>>>>>>>> @@ -505,6 +505,9 @@ pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
>>>>>>>> /* Install a pte for a particular vaddr in kernel space. */
>>>>>>>> void set_pte_vaddr(unsigned long vaddr, pte_t pte);
>>>>>>>> 
>>>>>>>> +struct mm_struct;
>>>>>>>> +void set_mm_pte_vaddr(struct mm_struct *mm, unsigned long vaddr, pte_t pte);
>>>>>>>> +
>>>>>>>> #ifdef CONFIG_X86_32
>>>>>>>> extern void native_pagetable_init(void);
>>>>>>>> #else
>>>>>>>> diff --git a/arch/x86/include/asm/text-patching.h b/arch/x86/include/asm/text-patching.h
>>>>>>>> index 2ecd34e2d46c..cb364ea5b19d 100644
>>>>>>>> --- a/arch/x86/include/asm/text-patching.h
>>>>>>>> +++ b/arch/x86/include/asm/text-patching.h
>>>>>>>> @@ -38,4 +38,6 @@ extern void *text_poke(void *addr, const void *opcode, size_t len);
>>>>>>>> extern int poke_int3_handler(struct pt_regs *regs);
>>>>>>>> extern void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler);
>>>>>>>> 
>>>>>>>> +extern struct mm_struct *patching_mm;
>>>>>>>> +
>>>>>>>> #endif /* _ASM_X86_TEXT_PATCHING_H */
>>>>>>>> diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
>>>>>>>> index a481763a3776..fd8a950b0d62 100644
>>>>>>>> --- a/arch/x86/kernel/alternative.c
>>>>>>>> +++ b/arch/x86/kernel/alternative.c
>>>>>>>> @@ -11,6 +11,7 @@
>>>>>>>> #include <linux/stop_machine.h>
>>>>>>>> #include <linux/slab.h>
>>>>>>>> #include <linux/kdebug.h>
>>>>>>>> +#include <linux/mmu_context.h>
>>>>>>>> #include <asm/text-patching.h>
>>>>>>>> #include <asm/alternative.h>
>>>>>>>> #include <asm/sections.h>
>>>>>>>> @@ -701,8 +702,36 @@ void *text_poke(void *addr, const void *opcode, size_t len)
>>>>>>>>             WARN_ON(!PageReserved(pages[0]));
>>>>>>>>             pages[1] = virt_to_page(addr + PAGE_SIZE);
>>>>>>>>     }
>>>>>>>> -       BUG_ON(!pages[0]);
>>>>>>>> +
>>>>>>>>     local_irq_save(flags);
>>>>>>>> +       BUG_ON(!pages[0]);
>>>>>>>> +
>>>>>>>> +       /*
>>>>>>>> +        * During initial boot, it is hard to initialize patching_mm due to
>>>>>>>> +        * dependencies in boot order.
>>>>>>>> +        */
>>>>>>>> +       if (patching_mm) {
>>>>>>>> +               pte_t pte;
>>>>>>>> +               temporary_mm_state_t prev;
>>>>>>>> +
>>>>>>>> +               prev = use_temporary_mm(patching_mm);
>>>>>>>> +               pte = mk_pte(pages[0], PAGE_KERNEL);
>>>>>>>> +               set_mm_pte_vaddr(patching_mm, TEXT_POKE_ADDR, pte);
>>>>>>>> +               pte = mk_pte(pages[1], PAGE_KERNEL);
>>>>>>>> +               set_mm_pte_vaddr(patching_mm, TEXT_POKE_ADDR + PAGE_SIZE, pte);
>>>>>>>> +
>>>>>>>> +               memcpy((void *)(TEXT_POKE_ADDR | ((unsigned long)addr & ~PAGE_MASK)),
>>>>>>>> +                      opcode, len);
>>>>>>>> +
>>>>>>>> +               set_mm_pte_vaddr(patching_mm, TEXT_POKE_ADDR, __pte(0));
>>>>>>>> +               set_mm_pte_vaddr(patching_mm, TEXT_POKE_ADDR + PAGE_SIZE, __pte(0));
>>>>>>>> +               local_flush_tlb();
>>>>>>> 
>>>>>>> Hmm.  This is stuff busted on SMP, and it's IMO more complicated than
>>>>>>> needed.  How about getting rid of all the weird TLB flushing stuff and
>>>>>>> instead putting the mapping at vaddr - __START_KERNEL_map or whatever
>>>>>>> it is?  You *might* need to flush_tlb_mm_range() on module unload, but
>>>>>>> that's it.
>>>>>> 
>>>>>> I don’t see what’s wrong in SMP, since this entire piece of code should be
>>>>>> running under text_mutex.
>>>>>> 
>>>>>> I don’t quite understand your proposal. I really don’t want to have any
>>>>>> chance in which the page-tables for the poked address is not preallocated.
>>>>>> 
>>>>>> It is more complicated than needed, and there are redundant TLB flushes. The
>>>>>> reason I preferred to do it this way, is in order not to use other functions
>>>>>> that take locks during the software page-walk and not to duplicate existing
>>>>>> code. Yet, duplication might be the way to go.
>>>>>> 
>>>>>>>> +               sync_core();
>>>>>>> 
>>>>>>> I can't think of any case where sync_core() is needed.  The mm switch
>>>>>>> serializes.
>>>>>> 
>>>>>> Good point!
>>>>>> 
>>>>>>> Also, is there any circumstance in which any of this is used before at
>>>>>>> least jump table init?  All the early stuff is text_poke_early(),
>>>>>>> right?
>>>>>> 
>>>>>> Not before jump_label_init. However, I did not manage to get rid of the two
>>>>>> code-patches in text_poke(), since text_poke is used relatively early by
>>>>>> x86_late_time_init(), and at this stage kmem_cache_alloc() - which is needed
>>>>>> to duplicate init_mm - still fails.
>>>>> 
>>>>> Another correction: the populate_extra_pte() is not needed.
>>>>> 
>>>>> Anyhow, if you want to do this whole thing differently, I obviously will not
>>>>> object, but I think it will end up more complicated.
>>>>> 
>>>>> I think I finally understood your comment about "vaddr -
>>>>> __START_KERNEL_map”. I did something like that before, and it is not
>>>>> super-simple. You need not only to conditionally flush the TLB, but also
>>>>> to synchronize the PUD/PMD on changes. Don’t forget that module memory
>>>>> is installed even when BPF programs are installed.
>>>>> 
>>>>> Let me know if you want me to submit cleaner patches or you want to carry on
>>>>> yourself.
>>>> 
>>>> I think your approach is a good start and should be good enough (with
>>>> cleanups) as a fix for the bug.  But I think your code has the same
>>>> bug that we have now!  You're reusing the same address on multiple
>>>> CPUs without flushing.  You can easily fix it by forcing a flush
>>>> before loading the mm, which should be as simple as adding
>>>> flush_tlb_mm() before you load the mm.  (It won't actually flush
>>>> anything by itself, since the mm isn't loaded, but it will update the
>>>> bookkeeping so that switch_mm_irqs_off() flushes the mm.)
>>> 
>>> What am I missing?
>>> 
>>> We have a lock (text_mutex) which prevents the use of the new page-table
>>> hierarchy on multiple CPUs. In addition we have __set_pte_vaddr() which does
>>> a local TLB flush before the lock is released and before IRQs are enabled.
>>> So how can the PTE be cached on multiple CPUs?
>>> 
>>> Yes, __set_pte_vaddr() is ugly and flushes too much. I’ll try to remove some
>>> redundant TLB flushes, but these flushes are already there.
>> 
>> I missed that set_pte_vaddr() contained a flush.  It's probably nicer
>> to use one of the APIs that doesn't imply a flush.  If you end up with
>> vm_insert_pfn() or similar, you can use can use copy_to_user() if
>> needed :)
>> 
>>>> Also, please at least get rid of TEXT_POKE_ADDR.  If you don't want to
>>>> do the vaddr - __START_KERNEL_map thing, then at least pick an address
>>>> in the low half of the address space such as 0 :)  Ideally you'd only
>>>> use this thing late enough that you could even use the normal
>>>> insert_pfn (or similar) API for it, but that doesn't really matter.
>>> 
>>> Perhaps the vaddr - __START_KERNEL_map actually makes sense. I was
>>> misunderstanding it (again) before, thinking you want me to use vaddr (and
>>> not the delta). I’ll give it a try.
>> 
>> It does prevent easy preallocation of the intermediate tables.
> 
> Would we have to change the mapping address? If so, how about using hash
> of vaddr, which allow us to map it in one pte?

We do need two PTEs, and I will reuse the same two PTEs in order not to
waste memory on page-tables. I will try to randomize their address to be on
the safe side.

Powered by blists - more mailing lists