On 7/12/18 5:29 PM, Ming Lei wrote:

On Thu, Jul 12, 2018 at 10:36 PM, Hannes Reinecke [off-list ref] wrote:

quoted

Hi Jens, Christoph,

we're currently hunting down a silent data corruption occurring due to
commit 72ecad22d9f1 ("block: support a full bio worth of IO for
simplified bdev direct-io").

While the whole thing is still hazy on the details, the one thing we've
found is that reverting that patch fixes the data corruption.

And looking closer, I've found this:

static ssize_t
blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
{
        int nr_pages;

        nr_pages = iov_iter_npages(iter, BIO_MAX_PAGES + 1);
        if (!nr_pages)
                return 0;
        if (is_sync_kiocb(iocb) && nr_pages <= BIO_MAX_PAGES)
                return __blkdev_direct_IO_simple(iocb, iter, nr_pages);

        return __blkdev_direct_IO(iocb, iter, min(nr_pages, BIO_MAX_PAGES));
}

When checking the call path
__blkdev_direct_IO()->bio_alloc_bioset()->bvec_alloc()
I found that bvec_alloc() will fail if nr_pages > BIO_MAX_PAGES.

So why is there the check for 'nr_pages <= BIO_MAX_PAGES' ?
It's not that we can handle it in __blkdev_direct_IO() ...

Thanks for any clarification.

Maybe you can try the following patch from Christoph to see if it makes a
difference:

https://marc.info/?l=linux-kernel&m=153013977816825&w=2

That's not a bad idea.

-- 
Jens Axboe