Re: [PATCH] btrfs: Fix csum_tree_block to avoid tripping on -Werror=array-bounds

[David Sterba]

On Wed, May 24, 2023 at 07:46:42AM +0800, Qu Wenruo wrote:
> 
> 
> On 2023/5/24 03:32, David Sterba wrote:
> > On Tue, May 23, 2023 at 03:33:22PM +0800, Qu Wenruo wrote:
> >>
> >>
> >> On 2023/5/23 15:09, pengfuyuan wrote:
> >>>
> >>> When compiling on a mips 64-bit machine we get these warnings:
> >>>
> >>>       In file included from ./arch/mips/include/asm/cacheflush.h:13,
> >>> 	             from ./include/linux/cacheflush.h:5,
> >>> 	             from ./include/linux/highmem.h:8,
> >>> 		     from ./include/linux/bvec.h:10,
> >>> 		     from ./include/linux/blk_types.h:10,
> >>>                        from ./include/linux/blkdev.h:9,
> >>> 	             from fs/btrfs/disk-io.c:7:
> >>>       fs/btrfs/disk-io.c: In function ‘csum_tree_block’:
> >>>       fs/btrfs/disk-io.c:100:34: error: array subscript 1 is above array bounds of ‘struct page *[1]’ [-Werror=array-bounds]
> >>>         100 |   kaddr = page_address(buf->pages[i]);
> >>>             |                        ~~~~~~~~~~^~~
> >>>       ./include/linux/mm.h:2135:48: note: in definition of macro ‘page_address’
> >>>        2135 | #define page_address(page) lowmem_page_address(page)
> >>>             |                                                ^~~~
> >>>       cc1: all warnings being treated as errors
> >>>
> >>> We can check if i overflows to solve the problem. However, this doesn't make
> >>> much sense, since i == 1 and num_pages == 1 doesn't execute the body of the loop.
> >>> In addition, i < num_pages can also ensure that buf->pages[i] will not cross
> >>> the boundary. Unfortunately, this doesn't help with the problem observed here:
> >>> gcc still complains.
> >>
> >> So still false alerts, thus this bug should mostly be reported to GCC.
> >>
> >>>
> >>> To fix this, start the loop at index 0 instead of 1. Also, a conditional was
> >>> added to skip the case where the index is 0, so that the loop iterations follow
> >>> the desired logic, and it makes all versions of gcc happy.
> >>>
> >>> Signed-off-by: pengfuyuan <pengfuyuan@xxxxxxxxxx>
> >>> ---
> >>>    fs/btrfs/disk-io.c | 10 +++++++---
> >>>    1 file changed, 7 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
> >>> index fbf9006c6234..8b05d556d747 100644
> >>> --- a/fs/btrfs/disk-io.c
> >>> +++ b/fs/btrfs/disk-io.c
> >>> @@ -96,9 +96,13 @@ static void csum_tree_block(struct extent_buffer *buf, u8 *result)
> >>>    	crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
> >>>    			    first_page_part - BTRFS_CSUM_SIZE);
> >>>
> >>> -	for (i = 1; i < num_pages; i++) {
> >>> -		kaddr = page_address(buf->pages[i]);
> >>> -		crypto_shash_update(shash, kaddr, PAGE_SIZE);
> >>> +	for (i = 0; i < num_pages; i++) {
> >>> +		struct page *p = buf->pages[i];
> >>> +
> >>> +		if (i != 0) {
> >>> +			kaddr = page_address(p);
> >>> +			crypto_shash_update(shash, kaddr, PAGE_SIZE);
> >>
> >> Unfortunately this damages the readability.
> >>
> >> If you really want to starts from page index 0, I don't think doing this
> >> is the correct way.
> >>
> >> Instead, you may take the chance to merge the first
> >> crypto_shahs_update() call, so the overall procedure looks like this:
> >>
> >> static void csum_tree_block()
> >> {
> >> 	for (int i = 0; i < num_pages; i++) {
> >> 		int page_off = whatever_to_calculate_the_offset;
> >> 		int page_len = whatever_to_calculate_the_lengh;
> >> 		char *kaddr = page_address(buf->pages[i]) + page_off;
> >>
> >> 		crypto_shash_update(shash, kaddr, page_len);
> >> 	}
> >> 	memset();
> >> 	crypto_shash_final();
> >> }
> >>
> >> Although even with such change, I'm still not sure if it's any better or
> >> worse, as most of the calculation can still be bulky.
> >
> > Yeah I think the calculations would have to be conditional or keeping
> > some state. I'd like to keep the structure of the first page and the
> > rest.
> 
> Yeah, there would be conditional checks, but it turns out to be simpler
> like the following:
> 
> 	int cur = BTRFS_CSUM_SIZE;
> 
> 	for (int i = 0; i < num_pages; i++) {
> 		int range_end = min(eb->len, (i + 1) << PAGE_SHIFT);
> 		int page_len = range_end - cur;
> 		int page_off = offset_in_page(cur);
> 
> 		cypto_shash_update();
> 		cur = range_end;
> 	}
> 
> The only conditional thing is the min() call, but I'm not sure if this
> is any more readable though...

And then comes some joker and says "why don't you just handle the
first page separately and then loop over full pages" :)

We could also put the whole loop under #if INLINE_EB_PAGES > 1. I've
checked that this type of iteration over the pages is only present in
this function so this should not become a pattern that would spread
elsewhere.