Re: [PATCH resend] slub: Add back check for free nonslab objects

From: Vlastimil Babka
Date: Wed Sep 29 2021 - 12:39:28 EST


On 9/28/21 17:43, Matthew Wilcox wrote:
> On Mon, Sep 27, 2021 at 03:53:47PM +0800, Kefeng Wang wrote:
>> On 2021/9/27 15:22, Vlastimil Babka wrote:
>> > On 9/27/21 04:15, Kefeng Wang wrote:
>> > > After commit ("f227f0faf63b slub: fix unreclaimable slab stat for bulk
>> > > free"), the check for free nonslab page is replaced by VM_BUG_ON_PAGE,
>> > > which only check with CONFIG_DEBUG_VM enabled, but this config may
>> > > impact performance, so it only for debug.
>> > >
>> > > Commit ("0937502af7c9 slub: Add check for kfree() of non slab objects.")
>> > > add the ability, which should be needed in any configs to catch the
>> > > invalid free, they even could be potential issue, eg, memory corruption,
>> > > use after free and double-free, so replace VM_BUG_ON_PAGE to WARN_ON, and
>> > > add dump_page() to help use to debug the issue.
>> > There are other situations in SLUB (such as with smaller allocations that
>> > don't go directly to page allocator) where use after free and double-free
>> > are undetected in non-debug configs, and it's expected that anyone debugging
>> > them will enable slub_debug or even DEBUG_VM. Why should this special case
>> > with nonslab pages be different?
>>
>> I want the check back in kfree, this one is used  widely in driver, and the
>> probability
>>
>> of problem occurred is bigger in driver, especially in some out of tree
>> drivers.
>
> Why would we want to improve life for out of tree drivers? Drivers should
> be in-tree. That's been the Linux Way for thirty years.

Yes, there's a reason we distinguish VM_BUG_ON/VM_WARN_ON and plain
BUG_ON/WARN_ON. Picking arbitrarily one VM_ variant check and making it
always-enabled makes little sense to me, and doing it because of out of tree
drivers is certainly not a convincing argument. Commit f227f0faf63b was
correct in making it VM_BUG_ON.

> I remain sceptical that dump_page() is actually useful for debugging
> drivers anyway. dump_stack(), I could see -- that'll tell you which
> driver called kfree() on a bogus pointer. But how does dump_page() help?
>