Re: TLB flushes on fixmap changes

From: Andy Lutomirski
Date: Mon Aug 27 2018 - 18:32:28 EST


On Mon, Aug 27, 2018 at 2:55 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
> at 1:16 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>
>> at 12:58 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>
>>> On Mon, Aug 27, 2018 at 12:43 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>>>> at 12:10 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>>>>
>>>>> at 11:58 AM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>>>>
>>>>>> On Mon, Aug 27, 2018 at 11:54 AM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>>>>>>>> On Mon, Aug 27, 2018 at 10:34 AM, Nadav Amit <nadav.amit@xxxxxxxxx> 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 donât see whatâs wrong in SMP, since this entire piece of code should be
>> running under text_mutex.
>>
>> I donât quite understand your proposal. I really donât 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_mapâ. 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. Donât 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.)

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.