Message-ID: <CALCETrWxwpr+Xx0mCK1HUkanmCDOSRbw50VmebgoAgeNaaPAKg@mail.gmail.com>
Date:   Mon, 27 Aug 2018 11:45:38 -0700
From:   Andy Lutomirski <luto@...nel.org>
To:     Nadav Amit <nadav.amit@...il.com>
Cc:     Masami Hiramatsu <mhiramat@...nel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Andy Lutomirski <luto@...nel.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

On Mon, Aug 27, 2018 at 10:34 AM, Nadav Amit <nadav.amit@...il.com> wrote:
> at 1:05 AM, Masami Hiramatsu <mhiramat@...nel.org> wrote:
>
>> On Sun, 26 Aug 2018 20:26:09 -0700
>> Nadav Amit <nadav.amit@...il.com> wrote:
>>
>>> at 8:03 PM, Masami Hiramatsu <mhiramat@...nel.org> wrote:
>>>
>>>> On Sun, 26 Aug 2018 11:09:58 +0200
>>>> Peter Zijlstra <peterz@...radead.org> wrote:
>>>>
>>>>> On Sat, Aug 25, 2018 at 09:21:22PM -0700, Andy Lutomirski wrote:
>>>>>> I just re-read text_poke().  It's, um, horrible.  Not only is the
>>>>>> implementation overcomplicated and probably buggy, but it's SLOOOOOW.
>>>>>> It's totally the wrong API -- poking one instruction at a time
>>>>>> basically can't be efficient on x86.  The API should either poke lots
>>>>>> of instructions at once or should be text_poke_begin(); ...;
>>>>>> text_poke_end();.
>>>>>
>>>>> I don't think anybody ever cared about performance here. Only
>>>>> correctness. That whole text_poke_bp() thing is entirely tricky.
>>>>
>>>> Agreed. Self modification is a special event.
>>>>
>>>>> FWIW, before text_poke_bp(), text_poke() would only be used from
>>>>> stop_machine, so all the other CPUs would be stuck busy-waiting with
>>>>> IRQs disabled. These days, yeah, that's lots more dodgy, but yes
>>>>> text_mutex should be serializing all that.
>>>>
>>>> I'm still not sure that speculative page-table walk can be done
>>>> over the mutex. Also, if the fixmap area is for aliasing
>>>> pages (which always mapped to memory), what kind of
>>>> security issue can happen?
>>>
>>> The PTE is accessible from other cores, so just as we assume for L1TF that
>>> the every addressable memory might be cached in L1, we should assume and
>>> PTE might be cached in the TLB when it is present.
>>
>> Ok, so other cores can accidentally cache the PTE in TLB, (and no way
>> to shoot down explicitly?)
>
> There is way (although current it does not). But it seems that the consensus
> is that it is better to avoid it being mapped at all in remote cores.
>
>>> Although the mapping is for an alias, there are a couple of issues here.
>>> First, this alias mapping is writable, so it might an attacker to change the
>>> kernel code (following another initial attack).
>>
>> Combined with some buffer overflow, correct? If the attacker already can
>> write a kernel data directly, he is in the kernel mode.
>
> Right.
>
>>
>>> Second, the alias mapping is
>>> never explicitly flushed. We may assume that once the original mapping is
>>> removed/changed, a full TLB flush would take place, but there is no
>>> guarantee it actually takes place.
>>
>> Hmm, would this means a full TLB flush will not flush alias mapping?
>> (or, the full TLB flush just doesn't work?)
>
> It will flush the alias mapping, but currently there is no such explicit
> flush.
>
>>>> Anyway, from the viewpoint of kprobes, either per-cpu fixmap or
>>>> changing CR3 sounds good to me. I think we don't even need per-cpu,
>>>> it can call a thread/function on a dedicated core (like the first
>>>> boot processor) and wait :) This may prevent leakage of pte change
>>>> to other cores.
>>>
>>> I implemented per-cpu fixmap, but I think that it makes more sense to take
>>> peterz approach and set an entry in the PGD level. Per-CPU fixmap either
>>> requires to pre-populate various levels in the page-table hierarchy, or
>>> conditionally synchronize whenever module memory is allocated, since they
>>> can share the same PGD, PUD & PMD. While usually the synchronization is not
>>> needed, the possibility that synchronization is needed complicates locking.
>>
>> Could you point which PeterZ approach you said? I guess it will be
>> make a clone of PGD and use it for local page mapping (as new mm).
>> If so, yes it sounds perfectly fine to me.
>
> The thread is too long. What I think is best is having a mapping in the PGD
> level. I’ll try to give it a shot, and see what I get.
>
>>> Anyhow, having fixed addresses for the fixmap can be used to circumvent
>>> KASLR.
>>
>> I think text_poke doesn't mind using random address :)
>>
>>> I don’t think a dedicated core is needed. Anyhow there is a lock
>>> (text_mutex), so use_mm() can be used after acquiring the mutex.
>>
>> Hmm, use_mm() said;
>>
>> /*
>> * use_mm
>> *      Makes the calling kernel thread take on the specified
>> *      mm context.
>> *      (Note: this routine is intended to be called only
>> *      from a kernel thread context)
>> */
>>
>> So maybe we need a dedicated kernel thread for safeness?
>
> Yes, it says so. But I am not sure it cannot be changed, at least for this
> specific use-case. Switching kernel threads just for patching seems to me as
> an overkill.
>
> Let me see if I can get something half-reasonable doing so...
>

I don't understand at all how a kernel thread helps.  The useful bit
is to have a dedicated mm, which would involve setting up an mm_struct
and mapping the kernel and module text, EFI-style, in the user portion
of the mm.  But, to do the text_poke(), we'd just use the mm *without
calling use_mm*.

In other words, the following sequence should be (almost) just fine:

typedef struct {
  struct mm_struct *prev;
} temporary_mm_state_t;

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);
}

void unuse_temporary_mm(temporary_mm_state_t prev)
{
    lockdep_assert_irqs_disabled();
    switch_mm_irqs_off(NULL, prev.prev, current);
}

The only thing wrong with this that I can see is that it interacts
poorly with perf.  But perf is *already* busted in this regard.  The
following (whitespace damaged, sorry) should fix it:

commit b62bff5a8406d252de752cfe75068d0b73b9cdf0
Author: Andy Lutomirski <luto@...nel.org>
Date:   Mon Aug 27 11:41:55 2018 -0700

    x86/nmi: Fix some races in NMI uaccess

    In NMI context, we might be in the middle of context switching or in
    the middle of switch_mm_irqs_off().  In either case, CR3 might not
    match current->mm, which could cause copy_from_user_nmi() and
    friends to read the wrong memory.

    Fix it by adding a new nmi_uaccess_okay() helper and checking it in
    copy_from_user_nmi() and in __copy_from_user_nmi()'s callers.

    Signed-off-by: Andy Lutomirski <luto@...nel.org>

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 5f4829f10129..dfb2f7c0d019 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -2465,7 +2465,7 @@ perf_callchain_user(struct
perf_callchain_entry_ctx *entry, struct pt_regs *regs

     perf_callchain_store(entry, regs->ip);

-    if (!current->mm)
+    if (!nmi_uaccess_okay())
         return;

     if (perf_callchain_user32(regs, entry))
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 89a73bc31622..b23b2625793b 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -230,6 +230,22 @@ struct tlb_state {
 };
 DECLARE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate);

+/*
+ * Blindly accessing user memory from NMI context can be dangerous
+ * if we're in the middle of switching the current user task or
+ * switching the loaded mm.  It can also be dangerous if we
+ * interrupted some kernel code that was temporarily using a
+ * different mm.
+ */
+static inline bool nmi_uaccess_okay(void)
+{
+    struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
+    struct mm_struct *current_mm = current->mm;
+
+    return current_mm && loaded_mm == current_mm &&
+        loaded_mm->pgd == __va(read_cr3_pa());
+}
+
 /* Initialize cr4 shadow for this CPU. */
 static inline void cr4_init_shadow(void)
 {
diff --git a/arch/x86/lib/usercopy.c b/arch/x86/lib/usercopy.c
index c8c6ad0d58b8..c5f758430be2 100644
--- a/arch/x86/lib/usercopy.c
+++ b/arch/x86/lib/usercopy.c
@@ -19,6 +19,9 @@ copy_from_user_nmi(void *to, const void __user
*from, unsigned long n)
     if (__range_not_ok(from, n, TASK_SIZE))
         return n;

+    if (!nmi_uaccess_okay())
+        return n;
+
     /*
      * Even though this function is typically called from NMI/IRQ context
      * disable pagefaults so that its behaviour is consistent even when
diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
index 457b281b9339..f4b41d5a93dd 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -345,6 +345,9 @@ void switch_mm_irqs_off(struct mm_struct *prev,
struct mm_struct *next,
          */
         trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
     } else {
+        /* Let NMI code know that CR3 may not match expectations. */
+        this_cpu_write(cpu_tlbstate.loaded_mm, NULL);
+
         /* The new ASID is already up to date. */
         load_new_mm_cr3(next->pgd, new_asid, false);

What do you all think?

Powered by blists - more mailing lists