Re: [dpdk-dev] [PATCH 2/2] kni: fix rtnl deadlocks and race conditions v4

[Elad Nachman <eladv6@gmail.com>]

Hi Dan,

Thanks for the information but you are addressing a different problem. The
problem discussed here is making ifconfig up or down while the DPDK
application is running.

Elad.

בתאריך יום ב׳, 1 במרץ 2021, 23:26, מאת Dan Gora ‏<dg@adax.com>:

> This is from my git commit fixing this:
>
> kni: separate releasing netdev from freeing KNI interface
>
>     Currently the rte_kni kernel driver suffers from a problem where
>     when the interface is released, it generates a callback to the DPDK
>     application to change the interface state to Down.  However, after the
>     DPDK application handles the callback and generates a response back to
>     the kernel, the rte_kni driver cannot wake the thread which is asleep
>     waiting for the response, because it is holding the kni_link_lock
>     semaphore and it has already removed the 'struct kni_dev' from the
>     list of interfaces to poll for responses.
>
>     This means that if the KNI interface is in the Up state when
>     rte_kni_release() is called, it will always sleep for three seconds
>     until kni_net_release gives up waiting for a response from the DPDK
>     application.
>
>     To fix this, we must separate the step to release the kernel network
>     interface from the steps to remove the KNI interface from the list
>     of interfaces to poll.
>
>     When the kernel network interface is removed with unregister_netdev(),
>     if the interface is up, it will generate a callback to mark the
>     interface down, which calls kni_net_release().  kni_net_release() will
>     block waiting for the DPDK application to call rte_kni_handle_request()
>     to handle the callback, but it also needs the thread in the KNI driver
>     (either the per-dev thread for multi-thread or the per-driver thread)
>     to call kni_net_poll_resp() in order to wake the thread sleeping in
>     kni_net_release (actually kni_net_process_request()).
>
>     So now, KNI interfaces should be removed as such:
>
>     1) The user calls rte_kni_release().  This only unregisters the
>     netdev in the kernel, but touches nothing else.  This allows all the
>     threads to run which are necessary to handle the callback into the
>     DPDK application to mark the interface down.
>
>     2) The user stops the thread running rte_kni_handle_request().
>     After rte_kni_release() has been called, there will be no more
>     callbacks for that interface so it is not necessary.  It cannot be
>     running at the same time that rte_kni_free() frees all of the FIFOs
>     and DPDK memory for that KNI interface.
>
>     3) The user calls rte_kni_free().  This performs the RTE_KNI_IOCTL_FREE
>     ioctl which calls kni_ioctl_free().  This function removes the struct
>     kni_dev from the list of interfaces to poll (and kills the per-dev
>     kthread, if configured for multi-thread), then frees the memory in
>     the FIFOs.
>
>     Signed-off-by: Dan Gora <dg@adax.com>
>
> I'm not sure that this is exactly the problem that you're seeing, but
> it sounds like it to me.
>
> thanks
> dan
>
> On Mon, Mar 1, 2021 at 5:27 PM Dan Gora <dg@adax.com> wrote:
> >
> > Hi All,
> >
> > Sorry to butt in on this, but I fixed this same issue about 3 years
> > ago in my application, but I was never able to get the changes
> > integrated and eventually just gave up trying.
> >
> > The rule with KNI is:
> > 1) The app should have a separate control thread per rte_kni which
> > just spins calling rte_kni_handle_request().  This ensures that other
> > threads calling rte_kni_XXX functions will always get a response.
> >
> > 2) In order to deal with lockups and timeouts when closing the device, I
> sent
> > patches which separated the closing process into two steps:
> > rte_kni_release() which would unregister the underlying netdev, then
> > rte_kni_free() which would free the KNI portions of the KNI device.
> > When rte_kni_release() is called the kernel netdev is unregistered and
> > a response is sent back to the application, the control thread calling
> > rte_kni_handle_request() is still running, so the application will
> > still get a response back from the kernel and not lock up, the
> > application then kills the control thread so that
> > rte_kni_handle_request() is not called again, then the application
> > calls rte_kni_free() which frees all of the FIFOs and closes the
> > device.
> >
> > If anyone is interested the patches are probably still floating around
> > patchwork.  If not you can check them out here:
> >
> > https://github.com/danielgora/dpdk.git
> >
> > thanks-
> > dan
> >
> > On Mon, Mar 1, 2021 at 5:10 AM Igor Ryzhov <iryzhov@nfware.com> wrote:
> > >
> > > Stephen,
> > >
> > > No, I don't have a better proposal, but I think it is not correct to
> change
> > > the behavior of KNI (making link down without a real response).
> > > Even though we know that communicating with userspace under rtnl_lock
> is a
> > > bad idea, it works as it is for many years already.
> > >
> > > Elad,
> > >
> > > I agree with you that KNI should be removed from the main tree if it
> is not
> > > possible to fix this __dev_close_many issue.
> > > There were discussions about this multiple times already, but no one is
> > > working on this AFAIK.
> > > Last time the discussion was a month ago:
> > > https://www.mail-archive.com/dev@dpdk.org/msg196033.html
> > >
> > > Igor
> > >
> > > On Fri, Feb 26, 2021 at 8:43 PM Elad Nachman <eladv6@gmail.com> wrote:
> > >
> > > > The way the kernel handles its locks and lists for the dev close many
> > > > path, there is no way you can go around this with rtnl unlocked :
> > > > "
> > > >
> > > > There is a race condition in __dev_close_many() processing the
> > > > close_list while the application terminates.
> > > > It looks like if two vEth devices are terminating,
> > > > and one releases the rtnl lock, the other takes it,
> > > > updating the close_list in an unstable state,
> > > > causing the close_list to become a circular linked list,
> > > > hence list_for_each_entry() will endlessly loop inside
> > > > __dev_close_many() .
> > > >
> > > > "
> > > > And I don't expect David Miller will bend the kernel networking for
> DPDK
> > > > or KNI.
> > > >
> > > > But - Stephen - if you can personally convince David to accept a
> > > > kernel patch which will separate the close_list locking mechanism to
> a
> > > > separate (RCU?) lock, then I can introduce first a patch to the
> kernel
> > > > which will add a lock for the close_list, this way rtnl can be
> > > > unlocked for the if down case.
> > > >
> > > > After that kernel patch, your original patch + relocation of the sync
> > > > mutex locking will do the job .
> > > >lockups
> > > > Otherwise, rtnl has to be kept locked all of the way for the if down
> > > > case in order to prevent corruption causing a circular linked list
> out
> > > > of the close_list, causing a hang in the kernel.
> > > >lockups
> > > > Currently, the rtnl lock is the only thing keeping the close_list
> from
> > > > corruption.
> > > >
> > > > If you doubt rtnl cannot be unlocked for dev close path, you can
> > > > consult David for his opinion, as I think it is critical to
> understand
> > > > what the kernel can or cannot do, or expects to be done before we can
> > > > unlock its locks as we wish inside rte_kni.ko .
> > > >
> > > > Otherwise, if we are still in disagreement on how to patch this set
> of
> > > > problems, I think the responsible way around it is to completely
> > > > remove kni from the main dpdk tree and move it to dpdk-kmods
> > > > repository.
> > > >
> > > > I know BSD style open-source does not carry legal responsibility from
> > > > the developers, but I think when a bunch of developers know a piece
> of
> > > > code is highly buggy, they should not leave it for countless new
> users
> > > > to bounce their head desperately against, if they cannot agree on a
> > > > correct way to solve the bunch of problems, of which I think we all
> > > > agree exist (we just do not agree on the proper solution or patch)...
> > > >
> > > > That's my two cents,
> > > >
> > > > Elad.
> > > >
> > > > On Fri, Feb 26, 2021 at 5:49 PM Stephen Hemminger
> > > > <stephen@networkplumber.org> wrote:
> > > > >
> > > > > On Fri, 26 Feb 2021 00:01:01 +0300
> > > > > Igor Ryzhov <iryzhov@nfware.com> wrote:
> > > > >
> > > > > > Hi Elad,
> > > > > >
> > > > > > Thanks for the patch, but this is still NACK from me.
> > > > > >
> > > > > > The only real advantage of KNI over other exceptional-path
> techniques
> > > > > > like virtio-user is the ability to configure DPDK-managed
> interfaces
> > > > > > directly
> > > > > > from the kernel using well-known utils like iproute2. A very
> important
> > > > part
> > > > > > of this is getting responses from the DPDK app and knowing the
> actual
> > > > > > result of command execution.
> > > > > > If you're making async requests to the application and you don't
> know
> > > > > > the result, then what's the point of using KNI at all?
> > > > > >
> > > > > > Igor
> > > > >
> > > > > Do you have a better proposal that keeps the request result but
> does not
> > > > > call userspace with lock held.
> > > > >
> > > > > PS: I also have strong dislike of KNI, as designed it would have
> been
> > > > rejected
> > > > > by Linux kernel developers.  A better solution would be userspace
> > > > version of
> > > > > something like devlink devices. But doing control operations by
> proxy is
> > > > > a locking nightmare.
> > > >
>
>
>
> --
> Dan Gora
> Software Engineer
>
> Adax, Inc.
> Rua Dona Maria Alves, 1070 Casa 5
> Centro
> Ubatuba, SP
> CEP 11680-000
> Brasil
>
> Tel: +55 (12) 3833-1021  (Brazil and outside of US)
>     : +1 (510) 859-4801  (Inside of US)
>     : dan_gora (Skype)
>
> email: dg@adax.com
>