Re: [PATCH V1 18/32] osdep: import MADV_DOEXEC

[Alex Williamson]

On Wed, 19 Aug 2020 17:52:26 -0400
Steven Sistare <steven.sistare@oracle.com> wrote:

> On 8/17/2020 10:42 PM, Alex Williamson wrote:
> > On Mon, 17 Aug 2020 15:44:03 -0600
> > Alex Williamson <alex.williamson@redhat.com> wrote:
> >   
> >> On Mon, 17 Aug 2020 17:20:57 -0400
> >> Steven Sistare <steven.sistare@oracle.com> wrote:
> >>  
> >>> On 8/17/2020 4:48 PM, Alex Williamson wrote:    
> >>>> On Mon, 17 Aug 2020 14:30:51 -0400
> >>>> Steven Sistare <steven.sistare@oracle.com> wrote:
> >>>>       
> >>>>> On 7/30/2020 11:14 AM, Steve Sistare wrote:      
> >>>>>> Anonymous memory segments used by the guest are preserved across a 
> >>>>>> re-exec
> >>>>>> of qemu, mapped at the same VA, via a proposed madvise(MADV_DOEXEC) 
> >>>>>> option
> >>>>>> in the Linux kernel. For the madvise patches, see:
> >>>>>>
> >>>>>> https://lore.kernel.org/lkml/1595869887-23307-1-git-send-email-anthony.yznaga@oracle.com/
> >>>>>>
> >>>>>> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
> >>>>>> ---
> >>>>>>  include/qemu/osdep.h | 7 +++++++
> >>>>>>  1 file changed, 7 insertions(+)        
> >>>>>
> >>>>> Hi Alex,
> >>>>>   The MADV_DOEXEC functionality, which is a pre-requisite for the 
> >>>>> entire qemu 
> >>>>> live update series, is getting a chilly reception on lkml.  We could 
> >>>>> instead 
> >>>>> create guest memory using memfd_create and preserve the fd across exec. 
> >>>>>  However, 
> >>>>> the subsequent mmap(fd) will return a different VA than was used 
> >>>>> previously, 
> >>>>> which  is a problem for memory that was registered with vfio, as the 
> >>>>> original VA 
> >>>>> is remembered in the kernel struct vfio_dma and used in various kernel 
> >>>>> functions, 
> >>>>> such as vfio_iommu_replay.
> >>>>>
> >>>>> To fix, we could provide a VFIO_IOMMU_REMAP_DMA ioctl taking iova, 
> >>>>> size, and
> >>>>> new_vaddr.  The implementation finds an exact match for (iova, size) 
> >>>>> and replaces 
> >>>>> vaddr with new_vaddr.  Flags cannot be changed.
> >>>>>
> >>>>> memfd_create plus VFIO_IOMMU_REMAP_DMA would replace MADV_DOEXEC.
> >>>>> vfio on any form of shared memory (shm, dax, etc) could also be 
> >>>>> preserved across
> >>>>> exec with shmat/mmap plus VFIO_IOMMU_REMAP_DMA.
> >>>>>
> >>>>> What do you think      
> >>>>
> >>>> Your new REMAP ioctl would have parameters identical to the MAP_DMA
> >>>> ioctl, so I think we should just use one of the flag bits on the
> >>>> existing MAP_DMA ioctl for this variant.      
> >>>
> >>> Sounds good.
> >>>     
> >>>> Reading through the discussion on the kernel side there seems to be
> >>>> some confusion around why vfio needs the vaddr beyond the user call to
> >>>> MAP_DMA though.  Originally this was used to test for virtually
> >>>> contiguous mappings for merging and splitting purposes.  This is
> >>>> defunct in the v2 interface, however the vaddr is now used largely for
> >>>> mdev devices.  If an mdev device is not backed by an IOMMU device and
> >>>> does not share a container with an IOMMU device, then a user MAP_DMA
> >>>> ioctl essentially just registers the translation within the vfio
> >>>> container.  The mdev vendor driver can then later either request pages
> >>>> to be pinned for device DMA or can perform copy_to/from_user() to
> >>>> simulate DMA via the CPU.
> >>>>
> >>>> Therefore I don't see that there's a simple re-architecture of the vfio
> >>>> IOMMU backend that could drop vaddr use.        
> >>>
> >>> Yes.  I did not explain on lkml as you do here (thanks), but I reached 
> >>> the 
> >>> same conclusion.
> >>>     
> >>>> I'm a bit concerned this new
> >>>> remap proposal also raises the question of how do we prevent userspace
> >>>> remapping vaddrs racing with asynchronous kernel use of the previous
> >>>> vaddrs.        
> >>>
> >>> Agreed.  After a quick glance at the code, holding iommu->lock during 
> >>> remap might be sufficient, but it needs more study.    
> >>
> >> Unless you're suggesting an extended hold of the lock across the entire
> >> re-exec of QEMU, that's only going to prevent a race between a remap
> >> and a vendor driver pin or access, the time between the previous vaddr
> >> becoming invalid and the remap is unprotected.  
> 
> OK.  What if we exclude mediated devices?  Its appears they are the only
> ones where the kernel may async'ly use the vaddr, via call chains ending in 
> vfio_iommu_type1_pin_pages or vfio_iommu_type1_dma_rw_chunk.
> 
> The other functions that use dma->vaddr are
>     vfio_dma_do_map 
>     vfio_pin_map_dma 
>     vfio_iommu_replay 
>     vfio_pin_pages_remote
> and they are all initiated via userland ioctl (if I have traced all the code 
> paths correctly).  Thus iommu->lock would protect them.
> 
> We would block live update in qemu if the config includes a mediated device.
> 
> VFIO_IOMMU_REMAP_DMA would return EINVAL if the container has a mediated 
> device.

That's not a solution I'd really be in favor of.  We're eliminating an
entire class of devices because they _might_ make use of these
interfaces, but anyone can add a vfio bus driver, even exposing the
same device API, and maybe make use of some of these interfaces in that
driver.  Maybe we'd even have reason to do it in vfio-pci if we had
reason to virtualize some aspect of a device.  I think we're setting
ourselves up for a very complicated support scenario if we just
arbitrarily decide to deny drivers using certain interfaces.


> >>>> Are we expecting guest drivers/agents to quiesce the device,
> >>>> or maybe relying on clearing bus-master, for PCI devices, to halt DMA?   
> >>>>    
> >>>
> >>> No.  We want to support any guest, and the guest is not aware that qemu
> >>> live update is occurring.
> >>>     
> >>>> The vfio migration interface we've developed does have a mechanism to
> >>>> stop a device, would we need to use this here?  If we do have a
> >>>> mechanism to quiesce the device, is the only reason we're not unmapping
> >>>> everything and remapping it into the new address space the latency in
> >>>> performing that operation?  Thanks,      
> >>>
> >>> Same answer - we don't require that the guest has vfio migration support. 
> >>>    
> >>
> >> QEMU toggling the runstate of the device via the vfio migration
> >> interface could be done transparently to the guest, but if your
> >> intention is to support any device (where none currently support the
> >> migration interface) perhaps it's a moot point.    
> 
> That sounds useful when devices support.  Can you give me some function names
> or references so I can study this qemu-based "vfio migration interface".

The uAPI is documented in commit a8a24f3f6e38.  We're still waiting on
the QEMU support or implementation in an mdev vendor driver.
Essentially migration exposes a new region of the device which would be
implemented by the vendor driver.  A register within that region
manipulates the device state, so a device could be stopped by clearing
the 'run' bit in that register.


> >> It seems like this
> >> scheme only works with IOMMU backed devices where the device can
> >> continue to operate against pinned pages, anything that might need to
> >> dynamically pin pages against the process vaddr as it's running async
> >> to the QEMU re-exec needs to be blocked or stopped.  Thanks,  
> 
> Yes, true of this remap proposal.
> 
> I wanted to unconditionally support all devices, which is why I think that
> 
> MADV_DOEXEC is a nifty solution.  If you agree, please add your voice to the
> 
> lkml discussion.
> 
> > Hmm, even if a device is running against pinned memory, how do we
> > handle device interrupts that occur during QEMU's downtime?  I see that
> > we reconfigure interrupts, but does QEMU need to drain the eventfd and
> > manually inject those missed interrupts or will setting up the irqfds
> > trigger a poll?  Thanks,  
> 
> My existing code is apparently deficient in this area; I close the pre-exec 
> eventfd,
> and post exec create a new eventfd and attach it to the vfio device.  Jason 
> and I
> are discussing alternatives in this thread of the series:
>   
> https://lore.kernel.org/qemu-devel/https0da862c8-74bc-bf06-a436-4ebfcb9dd8d4@oracle.com/
> 
> I am hoping that unconditionally injecting a (potentially spurious) interrupt 
> on a 
> new eventfd after exec will solve the problem.

That's sloppy, but maybe sufficient, but I agree with Jason's concern
about treading carefully around anything that would cause the interrupt
state of the device to be modified, which certainly might not be
transparent to the device.  Then there's the issue of what would happen
if a fatal AER event occurred while the eventfd is disconnected.  We
wouldn't want to generate a spurious event on that channel.  The device
request eventfd will retry from the kernel side, so simply reconnecting
it should work.  Each type of virtual interrupt will need to have a
plan for what to do around this disconnected period, and like the error
reporting one, it might not be safe to lose the interrupt nor inject a
spurious interrupt.

Regarding emulated state in QEMU, yes QEMU does write all config to
the kernel, where some things might be emulated in the kernel, but
there are also things emulated in QEMU.  See for example
vdev->emulated_config_bits.  Writing everything to the kernel is just a
simplification because we know the kernel will drop writes that it
doesn't allow.  It's essentially a catch-all.  Thanks,

Alex