[PATCH RESEND 1/1] um: oops on accessing a non-present page in the vmalloc area

[Petr Tesarik]

On 3/21/2024 5:44 AM, David Gow wrote:
> On Fri, 23 Feb 2024 at 22:07, Petr Tesarik <petrtesarik at huaweicloud.com> wrote:
>>
>> From: Petr Tesarik <petr.tesarik1 at huawei-partners.com>
>>
>> If a segmentation fault is caused by accessing an address in the vmalloc
>> area, check that the target page is present.
>>
>> Currently, if the kernel hits a guard page in the vmalloc area, UML blindly
>> assumes that the fault is caused by a stale mapping and will be fixed by
>> flush_tlb_kernel_vm(). Unsurprisingly, if the fault is caused by accessing
>> a guard page, no mapping is created, and when the faulting instruction is
>> restarted, it will cause exactly the same fault again, effectively creating
>> an infinite loop.
>>
>> Signed-off-by: Petr Tesarik <petr.tesarik1 at huawei-partners.com>
>> ---
>>  arch/um/kernel/trap.c | 4 ++++
>>  1 file changed, 4 insertions(+)
>>
>> diff --git a/arch/um/kernel/trap.c b/arch/um/kernel/trap.c
>> index 6d8ae86ae978..d5b85f1bfe33 100644
>> --- a/arch/um/kernel/trap.c
>> +++ b/arch/um/kernel/trap.c
>> @@ -206,11 +206,15 @@ unsigned long segv(struct faultinfo fi, unsigned long ip, int is_user,
>>         int err;
>>         int is_write = FAULT_WRITE(fi);
>>         unsigned long address = FAULT_ADDRESS(fi);
>> +       pte_t *pte;
>>
>>         if (!is_user && regs)
>>                 current->thread.segv_regs = container_of(regs, struct pt_regs, regs);
>>
>>         if (!is_user && (address >= start_vm) && (address < end_vm)) {
>> +               pte = virt_to_pte(&init_mm, address);
>> +               if (!pte_present(*pte))
>> +                       page_fault_oops(regs, address, ip);
> 
> page_fault_oops() appears to be private to arch/x86/mm/fault.c, so
> can't be used here?
> Also, it accepts struct pt_regs*, not struct uml_pt_regs*, so would
> need to at least handle the type difference here.

Argh, you're right. This was originally a two-patch series, but Richard
wanted improvements in the implementation which would require more
effort, see here:

http://lists.infradead.org/pipermail/linux-um/2024-January/006406.html

So I wanted to fix only the infinite loop, but in the mean time I forgot
about the dependency on the first patch:

http://lists.infradead.org/pipermail/linux-um/2023-December/006380.html

That's because a quick git grep page_fault_oops found the function. It
was my mistake that I did not notice the other page_fault_oops() earlier.

OK, please forget about this patch for now; I must rework it.

> Could we equally avoid the infinite loop here by putting the
> 'flush_tlb_kernel_vm();goto out;' behind a if (pte_present(...))
> check, and let the rest of the UML checks panic or oops if required.
> (Actually OOPSing where we can under UML would be nice to do at some
> point anyway, but is a bigger issue than just fixing a bug, IMO.)

Yes, that would be the best quick fix until I get to implementing all
the blows and whistles (oops_* helpers, notification chains, tainting,
etc.).

Petr T

> Or am I lacking a prerequisite patch or applying this to the wrong
> version (or otherwise missing something), as it definitely doesn't
> build here.
> 
> Cheers,
> -- David