Re: [PATCH] scatterlist: prevent invalid free when alloc fails

[Tejun Heo <tj@kernel.org>]

Hello,

First of all, the patch is line-wrapped.  Please check your email
settings.

On 08/26/2010 06:04 PM, Jeffrey Carlyle wrote:
> When alloc fails, free_table is being called. Depending on the number of
> bytes requested, we determine if we are going to call _get_free_page()
> or kmalloc(). When alloc fails, our math is wrong (due to sg_size - 1),
> and the last buffer is wrongfully assumed to have been allocated by
> kmalloc. Hence, kfree gets called and a panic occurs.
> 
> This fix sets the end marker and allows the proper freeing of the
> buffers.
> 
> Signed-off-by: Olusanya Soyannwo <c23746@motorola.com>
> Cc: Tejun Heo <tj@kernel.org>
> Cc: Jens Axboe <jaxboe@fusionio.com>
> Signed-off-by: Jeffrey Carlyle <jeff.carlyle@motorola.com>
> ---
>  lib/scatterlist.c |   11 +++++++++--
>  1 files changed, 9 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/scatterlist.c b/lib/scatterlist.c
> index a5ec428..acf2c6e 100644
> --- a/lib/scatterlist.c
> +++ b/lib/scatterlist.c
> @@ -163,7 +163,7 @@ void __sg_free_table(struct sg_table *table,
> unsigned int max_ents,
>  		return;
> 
>  	sgl = table->sgl;
> -	while (table->orig_nents) {
> +	while (table->orig_nents && sgl) {
>  		unsigned int alloc_size = table->orig_nents;
>  		unsigned int sg_size;

Why is this change necessary?

> @@ -227,6 +227,7 @@ int __sg_alloc_table(struct sg_table *table,
> unsigned int nents,
>  {
>  	struct scatterlist *sg, *prv;
>  	unsigned int left;
> +	unsigned int total_alloc = 0;
> 
>  #ifndef ARCH_HAS_SG_CHAIN
>  	BUG_ON(nents > max_ents);
> @@ -248,8 +249,14 @@ int __sg_alloc_table(struct sg_table *table,
> unsigned int nents,
>  		left -= sg_size;
> 
>  		sg = alloc_fn(alloc_size, gfp_mask);
> -		if (unlikely(!sg))
> +		if (unlikely(!sg)) {
> +			table->orig_nents = total_alloc;
> +			/* mark the end of previous entry */
> +			sg_mark_end(&prv[alloc_size - 1]);

prv[alloc_size - 1] is already marked as end by sg_init_table() during
the previous iteration.  Also, prv can be NULL at this point.  AFAICS,
the only thing necessary would be "if (prv) table->nents++", no?

Thanks.

-- 
tejun