RE: [PATCH 1/2] topology: use bin_attribute to avoid buff overflow

[Song Bao Hua (Barry Song)]

> -----Original Message-----
> From: Greg KH [mailto:gregkh@xxxxxxxxxxxxxxxxxxx]
> Sent: Wednesday, June 2, 2021 9:06 PM
> To: tiantao (H) <tiantao6@xxxxxxxxxxxxx>
> Cc: Andy Shevchenko <andy.shevchenko@xxxxxxxxx>; tiantao (H)
> <tiantao6@xxxxxxxxxxxxx>; Linux Kernel Mailing List
> <linux-kernel@xxxxxxxxxxxxxxx>; Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>;
> Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>; Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx>; Rafael J. Wysocki <rafael@xxxxxxxxxx>;
> Jonathan Cameron <jonathan.cameron@xxxxxxxxxx>
> Subject: Re: [PATCH 1/2] topology: use bin_attribute to avoid buff overflow
> 
> On Wed, Jun 02, 2021 at 05:00:16PM +0800, tiantao (H) wrote:
> >
> > 在 2021/6/2 16:48, Andy Shevchenko 写道:
> > > On Wed, Jun 2, 2021 at 9:45 AM tiantao (H) <tiantao6@xxxxxxxxxx> wrote:
> > > > 在 2021/6/2 14:18, Greg KH 写道:
> > > > > On Wed, Jun 02, 2021 at 02:14:49PM +0800, tiantao (H) wrote:
> > > > > > 在 2021/6/1 12:58, Greg KH 写道:
> > > > > > > On Tue, Jun 01, 2021 at 10:56:49AM +0800, Tian Tao wrote:
> > > ...
> > >
> > > > > > > >     /**
> > > > > > > > + * bitmap_print_to_buf - convert bitmap to list or hex format
> ASCII string
> > > > > > > > + * @list: indicates whether the bitmap must be list
> > > > > > > > + * @buf: page aligned buffer into which string is placed
> > > > > > > > + * @maskp: pointer to bitmap to convert
> > > > > > > > + * @nmaskbits: size of bitmap, in bits
> > > > > > > > + * @off: offset in buf
> > > > > > > > + * @count: count that already output
> > > > > > > > + *
> > > > > > > > + * the role of bitmap_print_to_buf and bitmap_print_to_pagebuf
> is
> > > > > > > > + * the same, the difference is that the second parameter of
> > > > > > > > + * bitmap_print_to_buf can be more than one pagesize.
> > > > > > > > + */
> > > > > > > > +int bitmap_print_to_buf(bool list, char *buf, const unsigned
> long *maskp,
> > > > > > > > +                  int nmaskbits, loff_t off, size_t count)
> > > > > > > > +{
> > > > > > > > +  int len, size;
> > > > > > > > +  void *data;
> > > > > > > > +  char *fmt = list ? "%*pbl\n" : "%*pb\n";
> > > > > > > > +
> > > > > > > > +  len = snprintf(NULL, 0, fmt, nmaskbits, maskp);
> > > > > > > > +
> > > > > > > > +  data = kvmalloc(len+1, GFP_KERNEL);
> > > > > > > > +  if (!data)
> > > > > > > > +          return -ENOMEM;
> > > > > > > > +
> > > > > > > > +  size = scnprintf(data, len+1, fmt, nmaskbits, maskp);
> > > > > > > > +  size = memory_read_from_buffer(buf, count, &off, data, size);
> > > > > > > > +  kvfree(data);
> > > > > > > > +
> > > > > > > > +  return size;
> > > > > > > Why is this so different from bitmap_print_to_pagebuf()?  Can't you
> just
> > > > > > > use this function as the "real" function and then change
> > > > > > > bitmap_print_to_pagebuf() to call it with a size of PAGE_SIZE?
> > > > > > Do you mean do following change, is that correct? :-)
> > > > > Maybe, it is whitespace corrupted, and it still feels like this function
> > > > > is much bigger than it needs to be given the function it is replacing
> is
> > > > > only a simple sprintf() call.
> > > > >
> > > > > > +int bitmap_print_to_buf(bool list, char *buf, const unsigned long
> *maskp,
> > > > > > +                       int nmaskbits, loff_t off, size_t count)
> > > > > > +{
> > > > > > +       int len, size;
> > > > > > +       void *data;
> > > > > > +       const char *fmt = list ? "%*pbl\n" : "%*pb\n";
> > > > > > +
> > > > > > +       if (off == LLONG_MAX && count == PAGE_SIZE - offset_in_page(buf))
> > > > > > +               return scnprintf(buf, count, fmt, nmaskbits, maskp);
> > > > > > +
> > > > > > +       len = snprintf(NULL, 0, fmt, nmaskbits, maskp);
> > > > > > +
> > > > > > +       data = kvmalloc(len+1, GFP_KERNEL);
> > > > > Why do you need to allocate more memory?  And why kvmalloc()?
> > > > Because the memory here will exceed a pagesize and we don't know the
> > > > exact size, we have to call
> > > >
> > > > snprintf first to get the actual size. kvmalloc() is used because when
> > > > physical memory is tight, kmalloc
> > > >
> > > > may fail, but vmalloc will succeed. It is not so bad that the memory is
> > > > not requested here.
> > > To me it sounds like the function is overengineered / lacks thought
> > > through / optimization.
> > > Can you provide a few examples that require the above algorithm?
> >
> > so you think we should use kmalloc instead of kvmalloc ?
> 
> What size bitmap would trigger a vmalloc() call to be forced here?
> 

According to kvmalloc_node(), only if size is larger than PAGE_SIZE,
kvmalloc will move to vmalloc if kmalloc fails to get memory. Otherwise,
it will return error.
void *kvmalloc_node(size_t size, gfp_t flags, int node)
{
	gfp_t kmalloc_flags = flags;
	void *ret;

	...

	ret = kmalloc_node(size, kmalloc_flags, node);

	/*
	 * It doesn't really make sense to fallback to vmalloc for sub page
	 * requests
	 */
	if (ret || size <= PAGE_SIZE)
		return ret;

	return __vmalloc_node(size, 1, flags, node,
			__builtin_return_address(0));
}

For bitmap, it is clear a large NR_CPUS can trigger vmalloc:
Code copy-paste from drivers/base/node.c:
	/* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
	BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));

But for list, it would be much more tricky. As a list could be as simple
as:
0-2047
It could also be as complex as:
0,1,3,5,7,9,11,13,.....,2045,2047

It totally depends on how the bitmap is like.

That's why tiantao's code is detecting size before malloc.

> thanks,
> 
> greg k-h

Thanks
Barry