Re: xen-netfront crash when detaching network while some network activity

["Marek Marczykowski-Górecki" <marmarek@invisiblethingslab.com>]

[-- Attachment #1: Type: text/plain, Size: 5548 bytes --]

On Wed, May 27, 2015 at 12:03:12AM +0200, Marek Marczykowski-Górecki wrote:
> On Tue, May 26, 2015 at 11:56:00AM +0100, David Vrabel wrote:
> > On 22/05/15 12:49, Marek Marczykowski-Górecki wrote:
> > > Hi all,
> > > 
> > > I'm experiencing xen-netfront crash when doing xl network-detach while
> > > some network activity is going on at the same time. It happens only when
> > > domU has more than one vcpu. Not sure if this matters, but the backend
> > > is in another domU (not dom0). I'm using Xen 4.2.2. It happens on kernel
> > > 3.9.4 and 4.1-rc1 as well.
> > > 
> > > Steps to reproduce:
> > > 1. Start the domU with some network interface
> > > 2. Call there 'ping -f some-IP'
> > > 3. Call 'xl network-detach NAME 0'
> > 
> > There's a use-after-free in xennet_remove().  Does this patch fix it?
> 
> Unfortunately not. Note that the crash is in xennet_disconnect_backend,
> which is called before xennet_destroy_queues in xennet_remove.
> I've tried to add napi_disable and even netif_napi_del just after
> napi_synchronize in xennet_disconnect_backend (which would probably
> cause crash when trying to cleanup the same later again), but it doesn't
> help - the crash is the same (still in gnttab_end_foreign_access called
> from xennet_disconnect_backend).

Finally I've found some more time to debug this... All tests redone on
v4.3-rc6 frontend and 3.18.17 backend.

Looking at xennet_tx_buf_gc(), I have an impression that shared page
(queue->grant_tx_page[id]) is/should be freed in some other means than
(indirectly) calling to free_page via gnttab_end_foreign_access. Maybe the bug
is that the page _is_ actually freed somewhere else already? At least changing
gnttab_end_foreign_access to gnttab_end_foreign_access_ref makes the crash
gone.

Relevant xennet_tx_buf_gc fragment:
            gnttab_end_foreign_access_ref(
                queue->grant_tx_ref[id], GNTMAP_readonly);
            gnttab_release_grant_reference(
                &queue->gref_tx_head, queue->grant_tx_ref[id]);
            queue->grant_tx_ref[id] = GRANT_INVALID_REF;
            queue->grant_tx_page[id] = NULL;
            add_id_to_freelist(&queue->tx_skb_freelist, queue->tx_skbs, id);
            dev_kfree_skb_irq(skb);

And similar fragment from xennet_release_tx_bufs:
        get_page(queue->grant_tx_page[i]);
        gnttab_end_foreign_access(queue->grant_tx_ref[i],
                      GNTMAP_readonly,
                      (unsigned long)page_address(queue->grant_tx_page[i]));
        queue->grant_tx_page[i] = NULL;
        queue->grant_tx_ref[i] = GRANT_INVALID_REF;
        add_id_to_freelist(&queue->tx_skb_freelist, queue->tx_skbs, i);
        dev_kfree_skb_irq(skb);

Note that both have dev_kfree_skb_irq, but the former use
gnttab_end_foreign_access_ref, while the later - gnttab_end_foreign_access.
Also note that the crash is in gnttab_end_foreign_access, so before
dev_kfree_skb_irq. If that would be double free, I'd expect crash in the later.

This change was introduced by cefe007 "xen-netfront: fix resource leak in
netfront". I'm not sure if changing gnttab_end_foreign_access back to
gnttab_end_foreign_access_ref would not (re)introduce some memory leak.

Let me paste again the error message:
[   73.718636] page:ffffea000043b1c0 count:0 mapcount:0 mapping:          (null) index:0x0
[   73.718661] flags: 0x3ffc0000008000(tail)
[   73.718684] page dumped because: VM_BUG_ON_PAGE(atomic_read(&page->_count) == 0)
[   73.718725] ------------[ cut here ]------------
[   73.718743] kernel BUG at include/linux/mm.h:338!

Also it all look quite strange - there is get_page() call just before
gnttab_end_foreign_access, but page->_count is still 0. Maybe it have something
to do how get_page() works on "tail" pages (whatever it means)?

    static inline void get_page(struct page *page)
    {
        if (unlikely(PageTail(page)))
            if (likely(__get_page_tail(page)))
                return;
        /*
         * Getting a normal page or the head of a compound page
         * requires to already have an elevated page->_count.
         */
        VM_BUG_ON_PAGE(atomic_read(&page->_count) <= 0, page);
        atomic_inc(&page->_count);
    }

which (I think) ends up in:

    static inline void __get_page_tail_foll(struct page *page,
                        bool get_page_head)
    {
        /*
         * If we're getting a tail page, the elevated page->_count is
         * required only in the head page and we will elevate the head
         * page->_count and tail page->_mapcount.
         *
         * We elevate page_tail->_mapcount for tail pages to force
         * page_tail->_count to be zero at all times to avoid getting
         * false positives from get_page_unless_zero() with
         * speculative page access (like in
         * page_cache_get_speculative()) on tail pages.
         */
        VM_BUG_ON_PAGE(atomic_read(&page->first_page->_count) <= 0, page);
        if (get_page_head)
            atomic_inc(&page->first_page->_count);
        get_huge_page_tail(page);
    }

So the use counter is incremented in page->first_page->_count, not
page->_count. But according to the comment, it should influence
page->_mapcount, but the error message says it does not.

Any ideas?

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?

[-- Attachment #2: Type: application/pgp-signature, Size: 473 bytes --]