[bpf,v3] net/xdp: fix zero-size allocation warning in xskq_create()

  Syzkaller reported the following issue:
 ------------[ cut here ]------------
 WARNING: CPU: 0 PID: 2807 at mm/vmalloc.c:3247 __vmalloc_node_range (mm/vmalloc.c:3361)
 Modules linked in:
 CPU: 0 PID: 2807 Comm: repro Not tainted 6.6.0-rc2+ #12
 Hardware name: Generic DT based system
 unwind_backtrace from show_stack (arch/arm/kernel/traps.c:258)
 show_stack from dump_stack_lvl (lib/dump_stack.c:107 (discriminator 1))
 dump_stack_lvl from __warn (kernel/panic.c:633 kernel/panic.c:680)
 __warn from warn_slowpath_fmt (./include/linux/context_tracking.h:153 kernel/panic.c:700)
 warn_slowpath_fmt from __vmalloc_node_range (mm/vmalloc.c:3361 (discriminator 3))
 __vmalloc_node_range from vmalloc_user (mm/vmalloc.c:3478)
 vmalloc_user from xskq_create (net/xdp/xsk_queue.c:40)
 xskq_create from xsk_setsockopt (net/xdp/xsk.c:953 net/xdp/xsk.c:1286)
 xsk_setsockopt from __sys_setsockopt (net/socket.c:2308)
 __sys_setsockopt from ret_fast_syscall (arch/arm/kernel/entry-common.S:68)

xskq_get_ring_size() uses struct_size() macro to safely calculate the
size of struct xsk_queue and q->nentries of desc members. But the
syzkaller repro was able to set q->nentries with the value initially
taken from copy_from_sockptr() high enough to return SIZE_MAX by
struct_size(). The next PAGE_ALIGN(size) is such case will overflow
the size_t value and set it to 0. This will trigger WARN_ON_ONCE in
vmalloc_user() -> __vmalloc_node_range().

The issue is reproducible on 32-bit arm kernel.

Reported-and-tested-by: syzbot+fae676d3cf469331fc89@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/000000000000c84b4705fb31741e@google.com/T/
Link: https://syzkaller.appspot.com/bug?extid=fae676d3cf469331fc89
Reported-by: syzbot+b132693e925cbbd89e26@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/000000000000e20df20606ebab4f@google.com/T/
Fixes: 9f78bf330a66 ("xsk: support use vaddr as ring")
Signed-off-by: Andrew Kanner <andrew.kanner@gmail.com>
---

Notes (akanner):
    v3:
      - free kzalloc-ed memory before return, the leak was noticed by
        Daniel Borkmann <daniel@iogearbox.net>
    v2: https://lore.kernel.org/all/20231002222939.1519-1-andrew.kanner@gmail.com/raw
      - use unlikely() optimization for the case with SIZE_MAX return from
        struct_size(), suggested by Alexander Lobakin
        <aleksander.lobakin@intel.com>
      - cc-ed 4 more maintainers, mentioned by cc_maintainers patchwork
        test
    
    v1: https://lore.kernel.org/all/20230928204440.543-1-andrew.kanner@gmail.com/T/
      - RFC notes:
        It was found that net/xdp/xsk.c:xsk_setsockopt() uses
        copy_from_sockptr() to get the number of entries (int) for cases
        with XDP_RX_RING / XDP_TX_RING and XDP_UMEM_FILL_RING /
        XDP_UMEM_COMPLETION_RING.
    
        Next in xsk_init_queue() there're 2 sanity checks (entries == 0)
        and (!is_power_of_2(entries)) for which -EINVAL will be returned.
    
        After that net/xdp/xsk_queue.c:xskq_create() will calculate the
        size multipling the number of entries (int) with the size of u64,
        at least.
    
        I wonder if there should be the upper bound (e.g. the 3rd sanity
        check inside xsk_init_queue()). It seems that without the upper
        limit it's quiet easy to overflow the allocated size (SIZE_MAX),
        especially for 32-bit architectures, for example arm nodes which
        were used by the syzkaller.
    
        In this patch I added a naive check for SIZE_MAX which helped to
        skip zero-size allocation after overflow, but maybe it's not quite
        right. Please, suggest if you have any thoughts about the
        appropriate limit for the size of these xdp rings.
    
        PS: the initial number of entries is 0x20000000 in syzkaller
        repro: syscall(__NR_setsockopt, (intptr_t)r[0], 0x11b, 3,
        0x20000040, 0x20);
    
        Link:
        https://syzkaller.appspot.com/text?tag=ReproC&x=10910f18280000

 net/xdp/xsk_queue.c | 5 +++++
 1 file changed, 5 insertions(+)
  

Hello:

This patch was applied to bpf/bpf.git (master)
by Martin KaFai Lau <martin.lau@kernel.org>:

On Thu,  5 Oct 2023 22:35:49 +0300 you wrote:
> Syzkaller reported the following issue:
>  ------------[ cut here ]------------
>  WARNING: CPU: 0 PID: 2807 at mm/vmalloc.c:3247 __vmalloc_node_range (mm/vmalloc.c:3361)
>  Modules linked in:
>  CPU: 0 PID: 2807 Comm: repro Not tainted 6.6.0-rc2+ #12
>  Hardware name: Generic DT based system
>  unwind_backtrace from show_stack (arch/arm/kernel/traps.c:258)
>  show_stack from dump_stack_lvl (lib/dump_stack.c:107 (discriminator 1))
>  dump_stack_lvl from __warn (kernel/panic.c:633 kernel/panic.c:680)
>  __warn from warn_slowpath_fmt (./include/linux/context_tracking.h:153 kernel/panic.c:700)
>  warn_slowpath_fmt from __vmalloc_node_range (mm/vmalloc.c:3361 (discriminator 3))
>  __vmalloc_node_range from vmalloc_user (mm/vmalloc.c:3478)
>  vmalloc_user from xskq_create (net/xdp/xsk_queue.c:40)
>  xskq_create from xsk_setsockopt (net/xdp/xsk.c:953 net/xdp/xsk.c:1286)
>  xsk_setsockopt from __sys_setsockopt (net/socket.c:2308)
>  __sys_setsockopt from ret_fast_syscall (arch/arm/kernel/entry-common.S:68)
> 
> [...]

Here is the summary with links:
  - [bpf,v3] net/xdp: fix zero-size allocation warning in xskq_create()
    https://git.kernel.org/bpf/bpf/c/90aeaa99f53e

You are awesome, thank you!
  

On 10/5/23 12:35 PM, Andrew Kanner wrote:
> Syzkaller reported the following issue:
>   ------------[ cut here ]------------
>   WARNING: CPU: 0 PID: 2807 at mm/vmalloc.c:3247 __vmalloc_node_range (mm/vmalloc.c:3361)
>   Modules linked in:
>   CPU: 0 PID: 2807 Comm: repro Not tainted 6.6.0-rc2+ #12
>   Hardware name: Generic DT based system
>   unwind_backtrace from show_stack (arch/arm/kernel/traps.c:258)
>   show_stack from dump_stack_lvl (lib/dump_stack.c:107 (discriminator 1))
>   dump_stack_lvl from __warn (kernel/panic.c:633 kernel/panic.c:680)
>   __warn from warn_slowpath_fmt (./include/linux/context_tracking.h:153 kernel/panic.c:700)
>   warn_slowpath_fmt from __vmalloc_node_range (mm/vmalloc.c:3361 (discriminator 3))
>   __vmalloc_node_range from vmalloc_user (mm/vmalloc.c:3478)
>   vmalloc_user from xskq_create (net/xdp/xsk_queue.c:40)
>   xskq_create from xsk_setsockopt (net/xdp/xsk.c:953 net/xdp/xsk.c:1286)
>   xsk_setsockopt from __sys_setsockopt (net/socket.c:2308)
>   __sys_setsockopt from ret_fast_syscall (arch/arm/kernel/entry-common.S:68)
> 
> xskq_get_ring_size() uses struct_size() macro to safely calculate the
> size of struct xsk_queue and q->nentries of desc members. But the
> syzkaller repro was able to set q->nentries with the value initially
> taken from copy_from_sockptr() high enough to return SIZE_MAX by
> struct_size(). The next PAGE_ALIGN(size) is such case will overflow
> the size_t value and set it to 0. This will trigger WARN_ON_ONCE in

Please ignore the pw-bot email. A question just came to my mind after applying.

> diff --git a/net/xdp/xsk_queue.c b/net/xdp/xsk_queue.c
> index f8905400ee07..c7e8bbb12752 100644
> --- a/net/xdp/xsk_queue.c
> +++ b/net/xdp/xsk_queue.c
> @@ -34,6 +34,11 @@ struct xsk_queue *xskq_create(u32 nentries, bool umem_queue)
>   	q->ring_mask = nentries - 1;
>   
>   	size = xskq_get_ring_size(q, umem_queue);
> +	if (unlikely(size == SIZE_MAX)) {

What if "size" is SIZE_MAX-1? Would it still overflow the PAGE_ALIGN below?

> +		kfree(q);
> +		return NULL;
> +	}
> +
>   	size = PAGE_ALIGN(size);
>   
>   	q->ring = vmalloc_user(size);
  

On Thu, Oct 05, 2023 at 06:00:46PM -0700, Martin KaFai Lau wrote:
[...]
> > diff --git a/net/xdp/xsk_queue.c b/net/xdp/xsk_queue.c
> > index f8905400ee07..c7e8bbb12752 100644
> > --- a/net/xdp/xsk_queue.c
> > +++ b/net/xdp/xsk_queue.c
> > @@ -34,6 +34,11 @@ struct xsk_queue *xskq_create(u32 nentries, bool umem_queue)
> >   	q->ring_mask = nentries - 1;
> >   	size = xskq_get_ring_size(q, umem_queue);
> > +	if (unlikely(size == SIZE_MAX)) {
> 
> What if "size" is SIZE_MAX-1? Would it still overflow the PAGE_ALIGN below?
> 
> > +		kfree(q);
> > +		return NULL;
> > +	}
> > +
> >   	size = PAGE_ALIGN(size);
> >   	q->ring = vmalloc_user(size);
> 

I asked myself the same question before v1. E.g. thinking about the
check: (size > SIZE_MAX - PAGE_SIZE + 1)

But xskq_create() is called after the check for
!is_power_of_2(entries) in xsk_init_queue(). So I tried the same
reproducer and divided the (nentries) value by 2 in a loop - it hits
either SIZE_MAX case or the normal cases without overflow (sometimes
throwing vmalloc error complaining about size which exceed total pages
in my arm setup).

So I can't see a way size will be SIZE_MAX-1, etc. Correct me if I'm
wrong, please.

PS: In the output below the first 2 values of (nentries) hit SIZE_MAX
case, the rest hit the normal case, vmalloc_user() is complaining
about 1 allocation:

0x20000000
0x10000000
0x8000000
[   41.759195][ T2807] pre PAGE_ALIGN size = 2147483968 (0x80000140), PAGE_SIZE = 4096 (0x1000)
[   41.759621][ T2807] repro-iter: vmalloc error: size 2147487744, exceeds total pages, mode:0xdc0(GFP_KERNEL|__GFP_ZERO), nodemask=(null),cpuset=/,mems_allowed=0
[...]
0x4000000
0x2000000
0x1000000
0x800000
0x400000
0x200000
0x100000
0x80000
0x40000
0x20000
0x10000
0x8000
0x4000
0x2000
0x1000
0x800
0x400
0x200
0x100
0x80
0x40
0x20
0x10
0x8
0x4
0x2
  

On 10/6/23 12:09 AM, Andrew Kanner wrote:
> On Thu, Oct 05, 2023 at 06:00:46PM -0700, Martin KaFai Lau wrote:
> [...]
>>> diff --git a/net/xdp/xsk_queue.c b/net/xdp/xsk_queue.c
>>> index f8905400ee07..c7e8bbb12752 100644
>>> --- a/net/xdp/xsk_queue.c
>>> +++ b/net/xdp/xsk_queue.c
>>> @@ -34,6 +34,11 @@ struct xsk_queue *xskq_create(u32 nentries, bool umem_queue)
>>>    	q->ring_mask = nentries - 1;
>>>    	size = xskq_get_ring_size(q, umem_queue);
>>> +	if (unlikely(size == SIZE_MAX)) {
>>
>> What if "size" is SIZE_MAX-1? Would it still overflow the PAGE_ALIGN below?
>>
>>> +		kfree(q);
>>> +		return NULL;
>>> +	}
>>> +
>>>    	size = PAGE_ALIGN(size);
>>>    	q->ring = vmalloc_user(size);
>>
> 
> I asked myself the same question before v1. E.g. thinking about the
> check: (size > SIZE_MAX - PAGE_SIZE + 1)
> 
> But xskq_create() is called after the check for
> !is_power_of_2(entries) in xsk_init_queue(). So I tried the same
> reproducer and divided the (nentries) value by 2 in a loop - it hits
> either SIZE_MAX case or the normal cases without overflow (sometimes
> throwing vmalloc error complaining about size which exceed total pages
> in my arm setup).
> 
> So I can't see a way size will be SIZE_MAX-1, etc. Correct me if I'm
> wrong, please.
> 
> PS: In the output below the first 2 values of (nentries) hit SIZE_MAX

Thanks for the explanation, so iiuc it means it will overflow the struct_size() 
first because of the is_power_of_2(nentries) requirement? Could you help adding 
some comment to explain? Thanks.

> case, the rest hit the normal case, vmalloc_user() is complaining
> about 1 allocation:
> 
> 0x20000000
> 0x10000000
> 0x8000000
> [   41.759195][ T2807] pre PAGE_ALIGN size = 2147483968 (0x80000140), PAGE_SIZE = 4096 (0x1000)
> [   41.759621][ T2807] repro-iter: vmalloc error: size 2147487744, exceeds total pages, mode:0xdc0(GFP_KERNEL|__GFP_ZERO), nodemask=(null),cpuset=/,mems_allowed=0
> [...]
> 0x4000000
> 0x2000000
> 0x1000000
> 0x800000
> 0x400000
> 0x200000
> 0x100000
> 0x80000
> 0x40000
> 0x20000
> 0x10000
> 0x8000
> 0x4000
> 0x2000
> 0x1000
> 0x800
> 0x400
> 0x200
> 0x100
> 0x80
> 0x40
> 0x20
> 0x10
> 0x8
> 0x4
> 0x2
>
  

On Fri, Oct 06, 2023 at 10:37:44AM -0700, Martin KaFai Lau wrote:
[...] 
> > > What if "size" is SIZE_MAX-1? Would it still overflow the PAGE_ALIGN below?
> > > 
> > > > +		kfree(q);
> > > > +		return NULL;
> > > > +	}
> > > > +
> > > >    	size = PAGE_ALIGN(size);
> > > >    	q->ring = vmalloc_user(size);
> > > 
> > 
> > I asked myself the same question before v1. E.g. thinking about the
> > check: (size > SIZE_MAX - PAGE_SIZE + 1)
> > 
> > But xskq_create() is called after the check for
> > !is_power_of_2(entries) in xsk_init_queue(). So I tried the same
> > reproducer and divided the (nentries) value by 2 in a loop - it hits
> > either SIZE_MAX case or the normal cases without overflow (sometimes
> > throwing vmalloc error complaining about size which exceed total pages
> > in my arm setup).
> > 
> > So I can't see a way size will be SIZE_MAX-1, etc. Correct me if I'm
> > wrong, please.
> > 
> > PS: In the output below the first 2 values of (nentries) hit SIZE_MAX
> 
> Thanks for the explanation, so iiuc it means it will overflow the
> struct_size() first because of the is_power_of_2(nentries) requirement?
> Could you help adding some comment to explain? Thanks.
>

The overflow happens because there's no upper limit for nentries
(userspace input). Let me add more context, e.g. from net/xdp/xsk.c:

static int xsk_setsockopt(struct socket *sock, int level, int optname,
                          sockptr_t optval, unsigned int optlen)
{
[...]
                if (copy_from_sockptr(&entries, optval, sizeof(entries)))
                        return -EFAULT;
[...]
                err = xsk_init_queue(entries, q, false);
[...]
}

'entries' is passed to xsk_init_queue() and there're 2 checks: for 0
and is_power_of_2() only, no upper bound check:

static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
                          bool umem_queue)
{
        struct xsk_queue *q;

        if (entries == 0 || *queue || !is_power_of_2(entries))
                return -EINVAL;

        q = xskq_create(entries, umem_queue);
        if (!q)
                return -ENOMEM;
[...]
}

The 'entries' value is next passed to struct_size() in
net/xdp/xsk_queue.c. If it's large enough - SIZE_MAX will be returned.

I'm not sure if some appropriate limit for the size of XDP_RX_RING /
XDP_TX_RING and XDP_UMEM_FILL_RING / XDP_UMEM_COMPLETION_RING rings
should be used. But anyway, vmalloc() will tell if it's not ok with
the requested allocation size.
  

On 10/6/23 4:24 PM, Andrew Kanner wrote:
>> Thanks for the explanation, so iiuc it means it will overflow the
>> struct_size() first because of the is_power_of_2(nentries) requirement?
>> Could you help adding some comment to explain? Thanks.
>>
> The overflow happens because there's no upper limit for nentries
> (userspace input). Let me add more context, e.g. from net/xdp/xsk.c:
> 
> static int xsk_setsockopt(struct socket *sock, int level, int optname,
>                            sockptr_t optval, unsigned int optlen)
> {
> [...]
>                  if (copy_from_sockptr(&entries, optval, sizeof(entries)))
>                          return -EFAULT;
> [...]
>                  err = xsk_init_queue(entries, q, false);
> [...]
> }
> 
> 'entries' is passed to xsk_init_queue() and there're 2 checks: for 0
> and is_power_of_2() only, no upper bound check:
> 
> static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
>                            bool umem_queue)
> {
>          struct xsk_queue *q;
> 
>          if (entries == 0 || *queue || !is_power_of_2(entries))
>                  return -EINVAL;
> 
>          q = xskq_create(entries, umem_queue);
>          if (!q)
>                  return -ENOMEM;
> [...]
> }
> 
> The 'entries' value is next passed to struct_size() in
> net/xdp/xsk_queue.c. If it's large enough - SIZE_MAX will be returned.

All make sense. I was mostly asking to add a comment at the "if (unlikely(size 
== SIZE_MAX)" check to explain this details on why checking SIZE_MAX is enough.
  

On Fri, Oct 06, 2023 at 04:58:18PM -0700, Martin KaFai Lau wrote:
> On 10/6/23 4:24 PM, Andrew Kanner wrote:
> > > Thanks for the explanation, so iiuc it means it will overflow the
> > > struct_size() first because of the is_power_of_2(nentries) requirement?
> > > Could you help adding some comment to explain? Thanks.
> > > 
> > The overflow happens because there's no upper limit for nentries
> > (userspace input). Let me add more context, e.g. from net/xdp/xsk.c:
> > 
> > static int xsk_setsockopt(struct socket *sock, int level, int optname,
> >                            sockptr_t optval, unsigned int optlen)
> > {
> > [...]
> >                  if (copy_from_sockptr(&entries, optval, sizeof(entries)))
> >                          return -EFAULT;
> > [...]
> >                  err = xsk_init_queue(entries, q, false);
> > [...]
> > }
> > 
> > 'entries' is passed to xsk_init_queue() and there're 2 checks: for 0
> > and is_power_of_2() only, no upper bound check:
> > 
> > static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
> >                            bool umem_queue)
> > {
> >          struct xsk_queue *q;
> > 
> >          if (entries == 0 || *queue || !is_power_of_2(entries))
> >                  return -EINVAL;
> > 
> >          q = xskq_create(entries, umem_queue);
> >          if (!q)
> >                  return -ENOMEM;
> > [...]
> > }
> > 
> > The 'entries' value is next passed to struct_size() in
> > net/xdp/xsk_queue.c. If it's large enough - SIZE_MAX will be returned.
> 
> All make sense. I was mostly asking to add a comment at the "if
> (unlikely(size == SIZE_MAX)" check to explain this details on why checking
> SIZE_MAX is enough.

Ok, I got it. Will add in v4.
Thanks.