Re: WARNING: kmalloc bug in input_mt_init_slots

[Dmitry Vyukov]

On Mon, Sep 24, 2018 at 8:41 PM, Dmitry Torokhov
<dmitry.torokhov@xxxxxxxxx> wrote:
> On Mon, Sep 24, 2018 at 03:55:04PM +0000, Christopher Lameter wrote:
>> On Mon, 24 Sep 2018, Dmitry Vyukov wrote:
>>
>> > On Mon, Sep 24, 2018 at 5:08 PM, Christopher Lameter <cl@xxxxxxxxx> wrote:
>> > > On Sun, 23 Sep 2018, Dmitry Vyukov wrote:
>> > >
>> > >> What was the motivation behind that WARNING about large allocations in
>> > >> kmalloc? Why do we want to know about them? Is the general policy that
>> > >> kmalloc calls with potentially large size requests need to use NOWARN?
>> > >> If this WARNING still considered useful? Or we should change it to
>> > >> pr_err?
>> > >
>> > > In general large allocs should be satisfied by the page allocator. The
>> > > slab allocators are used for allocating and managing small objects. The
>> > > page allocator has mechanisms to deal with large objects (compound pages,
>> > > multiple page sized allocs etc).
>> >
>> > I am asking more about the status of this warning. If it fires in
>> > input_mt_init_slots(), does it mean that input_mt_init_slots() needs
>> > to be fixed? If not, then we need to change this warning to something
>> > else.
>>
>> Hmmm.. kmalloc falls back to the page allocator already?
>>
>> See
>>
>> static __always_inline void *kmalloc(size_t size, gfp_t flags)
>> {
>>         if (__builtin_constant_p(size)) {
>
> It would not be a constant here though.
>
>>                 if (size > KMALLOC_MAX_CACHE_SIZE)
>>                         return kmalloc_large(size, flags);
>>
>>
>> Note that this uses KMALLOC_MAX_CACHE_SIZE which should be smaller than
>> KMALLOC_MAX_SIZE.
>>
>>
>> How large is the allocation? AFACIT nRequests larger than KMALLOC_MAX_SIZE
>> are larger than the maximum allowed by the page allocator. Thus the warning
>> and the NULL return.
>
> The size in this particular case is being derived from a value passed
> from userspace. Input core does not care about any limits on size of
> memory kmalloc() can support and is perfectly happy with getting NULL
> and telling userspace to go away with their silly requests by returning
> -ENOMEM.
>
> For the record: I definitely do not want to pre-sanitize size neither in
> uinput nor in input core.

Christopher,

Assuming that the size is large enough to fail in all allocators, is
this warning still useful? How? Should we remove it?