Re: [PATCH v22 2/5] ring-buffer: Introducing ring-buffer mapping functions

[David Hildenbrand <david@redhat.com>]

On 09.05.24 13:05, Vincent Donnefort wrote:
> On Tue, May 07, 2024 at 10:34:02PM -0400, Steven Rostedt wrote:
>> On Tue, 30 Apr 2024 12:13:51 +0100
>> Vincent Donnefort <vdonnefort@google.com> wrote:
>>
>>> +#ifdef CONFIG_MMU
>>> +static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
>>> +			struct vm_area_struct *vma)
>>> +{
>>> +	unsigned long nr_subbufs, nr_pages, vma_pages, pgoff = vma->vm_pgoff;
>>> +	unsigned int subbuf_pages, subbuf_order;
>>> +	struct page **pages;
>>> +	int p = 0, s = 0;
>>> +	int err;
>>> +
>>> +	/* Refuse MP_PRIVATE or writable mappings */
>>> +	if (vma->vm_flags & VM_WRITE || vma->vm_flags & VM_EXEC ||
>>> +	    !(vma->vm_flags & VM_MAYSHARE))
>>> +		return -EPERM;
>>> +
>>> +	/*
>>> +	 * Make sure the mapping cannot become writable later. Also tell the VM
>>> +	 * to not touch these pages (VM_DONTCOPY | VM_DONTEXPAND). Finally,
>>> +	 * prevent migration, GUP and dump (VM_IO).
>>> +	 */
>>> +	vm_flags_mod(vma, VM_DONTCOPY | VM_DONTEXPAND | VM_IO, VM_MAYWRITE);
>>
>> Do we really need the VM_IO?
>>
>> When testing this in gdb, I would get:
>>
>> (gdb) p tmap->map->subbuf_size
>> Cannot access memory at address 0x7ffff7fc2008
>>
>> It appears that you can't ptrace IO memory. When I removed that flag,
>> gdb has no problem reading that memory.
> 
> Yeah, VM_IO indeed implies DONTDUMP. VM_IO was part of Linus recommendations.

Yes, the VM should recognize that memory to some degree as being special 
already due to VM_MIXEDMAP and VM_DONTEXPAND.

#define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_PFNMAP | VM_MIXEDMAP)

So any of these flag achieve that (e.g., mlock_fixup() checks 
VM_SPECIAL). KSM similarly skips VM_DONTEXPAND and VM_MIXEDMAP (likely 
we should be using VM_SPECIAL in vma_ksm_compatible()). Not sure about 
page migration, likely its fine.

Thinking about MADV_DONTNEED, I can spot in 
madvise_dontneed_free_valid_vma() only that we disallow primarily VM_PFNMAP.

... I assume if user space MADV_DONTNEED's some pages we'll simply get a 
page fault later on access that will SIGBUS, handling that gracefully 
(we should double-check!).


> But perhaps, VM_DONTEXPAND and MIXEDMAP (implicitely set by vm_insert_pages) are
> enough protection?

Do we want to dump these pages? VM_DONTDUMP might be reasonabe then.

> 
> I don't see how anything could use GUP there and as David pointed-out on the
> previous version, it doesn't event prevent the GUP-fast path.

Yes, GUP-fast would still have worked under some conditions.

-- 
Cheers,

David / dhildenb