Message-ID: <20190329084941.785fae4d@x1.home>
Date:   Fri, 29 Mar 2019 08:49:41 -0600
From:   Alex Williamson <alex.williamson@...hat.com>
To:     Kirti Wankhede <kwankhede@...dia.com>
Cc:     Parav Pandit <parav@...lanox.com>,
        "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Neo Jia <cjia@...dia.com>
Subject: Re: [PATCH 8/8] vfio/mdev: Improve the create/remove sequence

On Thu, 28 Mar 2019 22:50:38 +0530
Kirti Wankhede <kwankhede@...dia.com> wrote:

> On 3/26/2019 9:00 PM, Parav Pandit wrote:
> > 
> >   
> >> -----Original Message-----
> >> From: Kirti Wankhede <kwankhede@...dia.com>
> >> Sent: Tuesday, March 26, 2019 2:06 AM
> >> To: Parav Pandit <parav@...lanox.com>; kvm@...r.kernel.org; linux-
> >> kernel@...r.kernel.org; alex.williamson@...hat.com
> >> Cc: Neo Jia <cjia@...dia.com>
> >> Subject: Re: [PATCH 8/8] vfio/mdev: Improve the create/remove sequence
> >>
> >>
> >>
> >> On 3/23/2019 4:50 AM, Parav Pandit wrote:  
> >>> There are five problems with current code structure.
> >>> 1. mdev device is placed on the mdev bus before it is created in the
> >>> vendor driver. Once a device is placed on the mdev bus without
> >>> creating its supporting underlying vendor device, an open() can get
> >>> triggered by userspace on partially initialized device.
> >>> Below ladder diagram highlight it.
> >>>
> >>>       cpu-0                                       cpu-1
> >>>       -----                                       -----
> >>>    create_store()
> >>>      mdev_create_device()
> >>>        device_register()
> >>>           ...
> >>>          vfio_mdev_probe()
> >>>          ...creates char device
> >>>                                         vfio_mdev_open()
> >>>                                           parent->ops->open(mdev)
> >>>                                             vfio_ap_mdev_open()
> >>>                                               matrix_mdev = NULL
> >>>         [...]
> >>>         parent->ops->create()
> >>>           vfio_ap_mdev_create()
> >>>             mdev_set_drvdata(mdev, matrix_mdev);
> >>>             /* Valid pointer set above */
> >>>  
> >>
> >> VFIO interface uses sysfs path of device or PCI device's BDF where it checks
> >> sysfs file for that device exist.
> >> In case of VFIO mdev device, above situation will never happen as open will
> >> only get called if sysfs entry for that device exist.
> >>
> >> If you don't use VFIO interface then this situation can arise. In that case
> >> probe() can be used for very basic initialization then create actual char
> >> device from create().
> >>  
> > I explained you that create() cannot do the heavy lifting work of creating netdev and rdma dev because at that stage driver doesn't know whether its getting used for VM or host.
> > create() needs to create the device that probe() can work on in stable manner.
> >   
> 
> You can identify if its getting used by VM or host from create(). Since
> probe() happens first, from create() you can check
> mdev_dev(mdev)->driver->name, if its 'vfio_mdev' then its getting used
> by VM, otherwise used by host.

If this is suggesting that we should have different create paths based
on driver name, please no.  Mdev devices should not be special, they're
attached to a bus which can host multiple drivers and devices on that
bus should have the ability to switch between drivers.  Not to mention
that a strcmp of a driver name to infer the purpose of a device is just
ugly as can be.

> >>> 2. Current creation sequence is,
> >>>    parent->ops_create()
> >>>    groups_register()
> >>>
> >>> Remove sequence is,
> >>>    parent->ops->remove()
> >>>    groups_unregister()
> >>> However, remove sequence should be exact mirror of creation sequence.
> >>> Once this is achieved, all users of the mdev will be terminated first
> >>> before removing underlying vendor device.
> >>> (Follow standard linux driver model).
> >>> At that point vendor's remove() ops shouldn't failed because device is
> >>> taken off the bus that should terminate the users.
> >>>  
> >>
> >> If VMM or user space application is using mdev device,
> >> parent->ops->remove() can return failure. In that case sysfs files
> >> shouldn't be removed. Hence above sequence is followed for remove.
> >>
> >> Standard linux driver model doesn't allow remove() to fail, but in of mdev
> >> framework, interface is defined to handle such error case.
> >>  
> > But the sequence is incorrect for wider use case.  
> >>  
> >>> 3. Additionally any new mdev driver that wants to work on mdev device
> >>> during probe() routine registered using mdev_register_driver() needs
> >>> to get stable mdev structure.
> >>>  
> >>
> >> Things that you are trying to handle with mdev structure from probe(),
> >> couldn't that be moved to create()?
> >>  
> > No, as explained before and above.
> > That approach just doesn't look right.
> >  
> 
> As I mentioned abouve, you can do that.

But it would be wrong to do so.

> >>  
> >>> 4. In following sequence, child devices created while removing mdev
> >>> parent device can be left out, or it may lead to race of removing half
> >>> initialized child mdev devices.
> >>>
> >>> issue-1:
> >>> --------
> >>>        cpu-0                         cpu-1
> >>>        -----                         -----
> >>>                                   mdev_unregister_device()
> >>>                                      device_for_each_child()
> >>>                                         mdev_device_remove_cb()
> >>>                                             mdev_device_remove()
> >>> create_store()
> >>>   mdev_device_create()                   [...]
> >>>        device_register()
> >>>                                   parent_remove_sysfs_files()
> >>>                                   /* BUG: device added by cpu-0
> >>>                                    * whose parent is getting removed.
> >>>                                    */
> >>>
> >>> issue-2:
> >>> --------
> >>>        cpu-0                         cpu-1
> >>>        -----                         -----
> >>> create_store()
> >>>   mdev_device_create()                   [...]
> >>>        device_register()
> >>>
> >>>        [...]                      mdev_unregister_device()
> >>>                                      device_for_each_child()
> >>>                                         mdev_device_remove_cb()
> >>>                                             mdev_device_remove()
> >>>
> >>>        mdev_create_sysfs_files()
> >>>        /* BUG: create is adding
> >>>         * sysfs files for a device
> >>>         * which is undergoing removal.
> >>>         */
> >>>                                  parent_remove_sysfs_files()
> >>>
> >>> 5. Below crash is observed when user initiated remove is in progress
> >>> and mdev_unregister_driver() completes parent unregistration.
> >>>
> >>>        cpu-0                         cpu-1
> >>>        -----                         -----
> >>> remove_store()
> >>>    mdev_device_remove()
> >>>    active = false;
> >>>                                   mdev_unregister_device()
> >>>                                     remove type
> >>>    [...]
> >>>    mdev_remove_ops() crashes.
> >>>
> >>> This is similar race like create() racing with mdev_unregister_device().
> >>>
> >>> mtty mtty: MDEV: Registered
> >>> iommu: Adding device 83b8f4f2-509f-382f-3c1e-e6bfe0fa1001 to group 57
> >>> vfio_mdev 83b8f4f2-509f-382f-3c1e-e6bfe0fa1001: MDEV: group_id = 57
> >>> mdev_device_remove sleep started mtty mtty: MDEV: Unregistering
> >>> mtty_dev: Unloaded!
> >>> BUG: unable to handle kernel paging request at ffffffffc027d668 PGD
> >>> af9818067 P4D af9818067 PUD af981a067 PMD 8583c3067 PTE 0
> >>> Oops: 0000 [#1] SMP PTI
> >>> CPU: 15 PID: 3517 Comm: bash Kdump: loaded Not tainted
> >>> 5.0.0-rc7-vdevbus+ #2 Hardware name: Supermicro
> >>> SYS-6028U-TR4+/X10DRU-i+, BIOS 2.0b 08/09/2016
> >>> RIP: 0010:mdev_device_remove_ops+0x1a/0x50 [mdev] Call Trace:
> >>>  mdev_device_remove+0xef/0x130 [mdev]
> >>>  remove_store+0x77/0xa0 [mdev]
> >>>  kernfs_fop_write+0x113/0x1a0
> >>>  __vfs_write+0x33/0x1b0
> >>>  ? rcu_read_lock_sched_held+0x64/0x70
> >>>  ? rcu_sync_lockdep_assert+0x2a/0x50
> >>>  ? __sb_start_write+0x121/0x1b0
> >>>  ? vfs_write+0x17c/0x1b0
> >>>  vfs_write+0xad/0x1b0
> >>>  ? trace_hardirqs_on_thunk+0x1a/0x1c
> >>>  ksys_write+0x55/0xc0
> >>>  do_syscall_64+0x5a/0x210
> >>>
> >>> Therefore, mdev core is improved in following ways to overcome above
> >>> issues.
> >>>
> >>> 1. Before placing mdev devices on the bus, perform vendor drivers
> >>> creation which supports the mdev creation.
> >>> This ensures that mdev specific all necessary fields are initialized
> >>> before a given mdev can be accessed by bus driver.
> >>>
> >>> 2. During remove flow, first remove the device from the bus. This
> >>> ensures that any bus specific devices and data is cleared.
> >>> Once device is taken of the mdev bus, perform remove() of mdev from
> >>> the vendor driver.
> >>>  
> >>
> >> If user space application is using the device and someone underneath
> >> remove the device from bus, how would use space application know that
> >> device is being removed?  
> > vfio_mdev guards and wait for device to get closed.
> > 
> > One sample trace is below.
> > [<0>] vfio_del_group_dev+0x34a/0x3c0 [vfio]
> > [<0>] mdev_remove+0x21/0x40 [mdev]
> > [<0>] device_release_driver_internal+0xe8/0x1b0
> > [<0>] bus_remove_device+0xf9/0x170
> > [<0>] device_del+0x168/0x350
> > [<0>] mdev_device_remove_common+0x1e/0x60 [mdev]
> > [<0>] mdev_device_remove_cb+0x1a/0x30 [mdev]
> > [<0>] device_for_each_child+0x47/0x90
> > [<0>] mdev_unregister_device+0xdb/0x100 [mdev]
> > [<0>] mtty_dev_exit+0x17/0x843 [mtty]
> > [<0>] __x64_sys_delete_module+0x16b/0x240
> > [<0>] do_syscall_64+0x5a/0x210
> > [<0>] entry_SYSCALL_64_after_hwframe+0x49/0xbe
> > [<0>] 0xffffffffffffffff
> >   
> >> If DMA is setup, user space application is accessing that memory and device
> >> is removed from bus - how will you restrict to not to remove that device? If
> >> remove() is not restricted then host might crash.
> >> I know Linux kernel device core model doesn't allow remove() to fail, but we
> >> had tackled that problem for mdev devices in this framework. I prefer not to
> >> change this behavior. This will regress existing working drivers.
> >>  
> > vfio layer ensures that open device cannot be removed from above trace.
> > 
> > Other drivers will follow similar method. In case of mlx5 driver which binds
> > to mdev follows standard driver model to terminate for this mdev device,
> > similar way for pci device.
> >   
> 
> But then remove() or write on 'remove' sysfs would block, which could be
> indefinite. For example in case of VM, it will block until VM is not
> shutdown.
> With current approach, write on 'remove' sysfs doesn't block.

OTOH, why should mdev be different than any other driver?  Blocking is
the current solution for all directly assigned vfio devices.  This is a
compromise between the device model not allowing an error return and
lack of support to be able to revoke mmaps to the device.  We already
have an interface in vfio to request a device from a cooperative user
(Maxim proposed adding this to the mdev interface), lacking a revoke
interface, that can be further escalated to killing the process.  What
we've heard previously when pursuing an error path from removing a
device is that all responsibility lies with the admin in using these
interfaces.  If a remove is requested, it should be honored.  If that
results in killing a task, the fault is on the admin.  Mdev is not its
own island to decide a different model. Thanks,

Alex

Powered by blists - more mailing lists