[Regression?] iptables broken on 32bit with pre-4.7-rc

[Regression?] iptables broken on 32bit with pre-4.7-rc

[John Stultz]

Hey Florian, Pablo,

In updating a 32bit arm device from 4.6 to Linus' current HEAD, I
noticed I was having some trouble with networking, and realized that
/proc/net/ip_tables_names was suddenly empty.

Digging through the registration process, it seems we're catching on the:

       if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
           target_offset + sizeof(struct xt_standard_target) != next_offset)
               return -EINVAL;

check added in 7ed2abddd20cf ("netfilter: x_tables: check standard
target size too").

Where next_offset seems to be 4 bytes larger then the the offset +
standard_target struct size.

Commenting out those checks (the commit doesn't revert cleanly), seems
to get things going again for me.

I'm not exactly sure how the next_offset value is set, so I'm hoping
the proper fix is more obvious to one of you.

thanks
-john

Re: [Regression?] iptables broken on 32bit with pre-4.7-rc

[Florian Westphal]

John Stultz <john.stultz@linaro.org> wrote:
> In updating a 32bit arm device from 4.6 to Linus' current HEAD, I
> noticed I was having some trouble with networking, and realized that
> /proc/net/ip_tables_names was suddenly empty.
> 
> Digging through the registration process, it seems we're catching on the:
> 
>        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
>            target_offset + sizeof(struct xt_standard_target) != next_offset)
>                return -EINVAL;
> 
> check added in 7ed2abddd20cf ("netfilter: x_tables: check standard
> target size too").
> 
> Where next_offset seems to be 4 bytes larger then the the offset +
> standard_target struct size.

I guess its because arm32 needs 8 byte alignment for 64bit
quantities.  So we can fix this either via XT_ALIGN()'ing the
target_offset + sizeof() result or by weakening the test to a '>'.

Since we already test proper alignment of start-of-rule in
check_entry_size_and_hooks() I'd suggest we just change the test
to fail only if the next offset is within the min size, i.e.:

diff --git a/net/netfilter/x_tables.c b/net/netfilter/x_tables.c
index c69c892..9643047 100644
--- a/net/netfilter/x_tables.c
+++ b/net/netfilter/x_tables.c
@@ -612,7 +612,7 @@ int xt_compat_check_entry_offsets(const void *base, const char *elems,
                return -EINVAL;
 
        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
-           target_offset + sizeof(struct compat_xt_standard_target) != next_offset)
+           target_offset + sizeof(struct compat_xt_standard_target) > next_offset)
                return -EINVAL;
 
        /* compat_xt_entry match has less strict aligment requirements,
@@ -694,7 +694,7 @@ int xt_check_entry_offsets(const void *base,
                return -EINVAL;
 
        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
-           target_offset + sizeof(struct xt_standard_target) != next_offset)
+           target_offset + sizeof(struct xt_standard_target) > next_offset)
                return -EINVAL;
 
        return xt_check_entry_match(elems, base + target_offset,

> I'm not exactly sure how the next_offset value is set, so I'm hoping
> the proper fix is more obvious to one of you.

Its the start of the next rule so it has to be properly aligned
via XT_ALIGN().  Only 32bit system I tested was plain x86 which
only needs 4byte alignment for u64...

Alternative would be something like this:

diff --git a/net/netfilter/x_tables.c b/net/netfilter/x_tables.c
index c69c892..ca16c26 100644
--- a/net/netfilter/x_tables.c
+++ b/net/netfilter/x_tables.c
@@ -612,7 +612,7 @@ int xt_compat_check_entry_offsets(const void *base, const char *elems,
                return -EINVAL;
 
        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
-           target_offset + sizeof(struct compat_xt_standard_target) != next_offset)
+           XT_COMPAT_ALIGN(target_offset + sizeof(struct compat_xt_standard_target)) != next_offset)
                return -EINVAL;
 
        /* compat_xt_entry match has less strict aligment requirements,
@@ -694,7 +694,7 @@ int xt_check_entry_offsets(const void *base,
                return -EINVAL;
 
        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
-           target_offset + sizeof(struct xt_standard_target) != next_offset)
+           XT_ALIGN(target_offset + sizeof(struct xt_standard_target)) != next_offset)
                return -EINVAL;
 
        return xt_check_entry_match(elems, base + target_offset,


but afaics the stricter check does not buy anything.

Re: [Regression?] iptables broken on 32bit with pre-4.7-rc

[John Stultz]

On Thu, May 26, 2016 at 2:51 AM, Florian Westphal <fw@strlen.de> wrote:
> John Stultz <john.stultz@linaro.org> wrote:
>> In updating a 32bit arm device from 4.6 to Linus' current HEAD, I
>> noticed I was having some trouble with networking, and realized that
>> /proc/net/ip_tables_names was suddenly empty.
>>
>> Digging through the registration process, it seems we're catching on the:
>>
>>        if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
>>            target_offset + sizeof(struct xt_standard_target) != next_offset)
>>                return -EINVAL;
>>
>> check added in 7ed2abddd20cf ("netfilter: x_tables: check standard
>> target size too").
>>
>> Where next_offset seems to be 4 bytes larger then the the offset +
>> standard_target struct size.
>
> I guess its because arm32 needs 8 byte alignment for 64bit
> quantities.  So we can fix this either via XT_ALIGN()'ing the
> target_offset + sizeof() result or by weakening the test to a '>'.
>
> Since we already test proper alignment of start-of-rule in
> check_entry_size_and_hooks() I'd suggest we just change the test
> to fail only if the next offset is within the min size, i.e.:
>
> diff --git a/net/netfilter/x_tables.c b/net/netfilter/x_tables.c
> index c69c892..9643047 100644
> --- a/net/netfilter/x_tables.c
> +++ b/net/netfilter/x_tables.c
> @@ -612,7 +612,7 @@ int xt_compat_check_entry_offsets(const void *base, const char *elems,
>                 return -EINVAL;
>
>         if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
> -           target_offset + sizeof(struct compat_xt_standard_target) != next_offset)
> +           target_offset + sizeof(struct compat_xt_standard_target) > next_offset)
>                 return -EINVAL;
>
>         /* compat_xt_entry match has less strict aligment requirements,
> @@ -694,7 +694,7 @@ int xt_check_entry_offsets(const void *base,
>                 return -EINVAL;
>
>         if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
> -           target_offset + sizeof(struct xt_standard_target) != next_offset)
> +           target_offset + sizeof(struct xt_standard_target) > next_offset)
>                 return -EINVAL;
>
>         return xt_check_entry_match(elems, base + target_offset,
>
>> I'm not exactly sure how the next_offset value is set, so I'm hoping
>> the proper fix is more obvious to one of you.
>
> Its the start of the next rule so it has to be properly aligned
> via XT_ALIGN().  Only 32bit system I tested was plain x86 which
> only needs 4byte alignment for u64...
>
> Alternative would be something like this:
>
> diff --git a/net/netfilter/x_tables.c b/net/netfilter/x_tables.c
> index c69c892..ca16c26 100644
> --- a/net/netfilter/x_tables.c
> +++ b/net/netfilter/x_tables.c
> @@ -612,7 +612,7 @@ int xt_compat_check_entry_offsets(const void *base, const char *elems,
>                 return -EINVAL;
>
>         if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
> -           target_offset + sizeof(struct compat_xt_standard_target) != next_offset)
> +           XT_COMPAT_ALIGN(target_offset + sizeof(struct compat_xt_standard_target)) != next_offset)
>                 return -EINVAL;
>
>         /* compat_xt_entry match has less strict aligment requirements,
> @@ -694,7 +694,7 @@ int xt_check_entry_offsets(const void *base,
>                 return -EINVAL;
>
>         if (strcmp(t->u.user.name, XT_STANDARD_TARGET) == 0 &&
> -           target_offset + sizeof(struct xt_standard_target) != next_offset)
> +           XT_ALIGN(target_offset + sizeof(struct xt_standard_target)) != next_offset)
>                 return -EINVAL;
>
>         return xt_check_entry_match(elems, base + target_offset,
>
>
> but afaics the stricter check does not buy anything.

I can validate that either of these solutions solves the issue for me.
But I'll leave the pick of which to merge up to you. :)

Thanks so much for the quick response!
-john