On Wed, Dec 29, 2021 at 05:32:16PM +0800, Jiasheng Jiang wrote:
> And then it will cause the BUG_ON() in sg_assign_page().
That's not true. The two BUG_ON()'s in sg_assign_page() are:
BUG_ON((unsigned long) page & (SG_CHAIN | SG_END));
#ifdef CONFIG_DEBUG_SG
BUG_ON(sg_is_chain(sg));
#endif
Neither will trigger due to page == NULL.
> do {
> + struct page *pg;
> unsigned int i = sgl->cur;
>
> plen = min_t(size_t, len, PAGE_SIZE);
>
> - sg_assign_page(sg + i, alloc_page(GFP_KERNEL));
> + pg = alloc_page(GFP_KERNEL);
> + if (!pg) {
> + err = -ENOMEM;
> + goto unlock;
> + }
> +
> + sg_assign_page(sg + i, pg);
> if (!sg_page(sg + i)) {
> err = -ENOMEM;
> goto unlock;
The NULL check is already done, just below the redundant one you're adding.
I do think it would be more logical to do the NULL check before
sg_assign_page(). So you could send a patch that moves it there. But it would
be a cleanup, not a fix.
- Eric
