On Wed, Jun 14, 2023 at 01:46:31PM +0200, Hannes Reinecke wrote:
> static void brd_free_pages(struct brd_device *brd)
> {
> - unsigned long pos = 0;
> - struct page *pages[FREE_BATCH];
> - int nr_pages;
> -
> - do {
> - int i;
> -
> - nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
> - (void **)pages, pos, FREE_BATCH);
> -
> - for (i = 0; i < nr_pages; i++) {
> - void *ret;
> -
> - BUG_ON(pages[i]->index < pos);
> - pos = pages[i]->index;
> - ret = radix_tree_delete(&brd->brd_pages, pos);
> - BUG_ON(!ret || ret != pages[i]);
> - __free_page(pages[i]);
> - }
> -
> - pos++;
> + struct page *page;
> + pgoff_t idx;
>
> - /*
> - * It takes 3.4 seconds to remove 80GiB ramdisk.
> - * So, we need cond_resched to avoid stalling the CPU.
> - */
> - cond_resched();
> + xa_for_each(&brd->brd_pages, idx, page) {
> + __free_page(page);
> + cond_resched_rcu();
This should be a regular cond_resched(). The body of the loop is run
without the RCU read lock held. Surprised none of the bots have noticed
an unlock-underflow. Perhaps they don't test brd ;-)
With that fixed,
Reviewed-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
