Re: [Qemu-devel] [PATCH v4 01/13] vfio: KABI for migration interface

[Alex Williamson <alex.williamson@redhat.com>]

On Sat, 22 Jun 2019 01:05:48 +0530
Kirti Wankhede <kwankhede@nvidia.com> wrote:

> On 6/21/2019 8:33 PM, Alex Williamson wrote:
> > On Fri, 21 Jun 2019 11:22:15 +0530
> > Kirti Wankhede <kwankhede@nvidia.com> wrote:
> >   
> >> On 6/20/2019 10:48 PM, Alex Williamson wrote:  
> >>> On Thu, 20 Jun 2019 20:07:29 +0530
> >>> Kirti Wankhede <kwankhede@nvidia.com> wrote:
> >>>     
> >>>> - Defined MIGRATION region type and sub-type.
> >>>> - Used 3 bits to define VFIO device states.
> >>>>     Bit 0 => _RUNNING
> >>>>     Bit 1 => _SAVING
> >>>>     Bit 2 => _RESUMING
> >>>>     Combination of these bits defines VFIO device's state during migration
> >>>>     _STOPPED => All bits 0 indicates VFIO device stopped.
> >>>>     _RUNNING => Normal VFIO device running state.
> >>>>     _SAVING | _RUNNING => vCPUs are running, VFIO device is running but start
> >>>>                           saving state of device i.e. pre-copy state
> >>>>     _SAVING  => vCPUs are stoppped, VFIO device should be stopped, and
> >>>>                           save device state,i.e. stop-n-copy state
> >>>>     _RESUMING => VFIO device resuming state.
> >>>>     _SAVING | _RESUMING => Invalid state if _SAVING and _RESUMING bits are set
> >>>> - Defined vfio_device_migration_info structure which will be placed at 0th
> >>>>   offset of migration region to get/set VFIO device related information.
> >>>>   Defined members of structure and usage on read/write access:
> >>>>     * device_state: (read/write)
> >>>>         To convey VFIO device state to be transitioned to. Only 3 bits are used
> >>>>         as of now.
> >>>>     * pending bytes: (read only)
> >>>>         To get pending bytes yet to be migrated for VFIO device.
> >>>>     * data_offset: (read only)
> >>>>         To get data offset in migration from where data exist during _SAVING
> >>>>         and from where data should be written by user space application during
> >>>>          _RESUMING state
> >>>>     * data_size: (read/write)
> >>>>         To get and set size of data copied in migration region during _SAVING
> >>>>         and _RESUMING state.
> >>>>     * start_pfn, page_size, total_pfns: (write only)
> >>>>         To get bitmap of dirty pages from vendor driver from given
> >>>>         start address for total_pfns.
> >>>>     * copied_pfns: (read only)
> >>>>         To get number of pfns bitmap copied in migration region.
> >>>>         Vendor driver should copy the bitmap with bits set only for
> >>>>         pages to be marked dirty in migration region. Vendor driver
> >>>>         should return 0 if there are 0 pages dirty in requested
> >>>>         range. Vendor driver should return -1 to mark all pages in the section
> >>>>         as dirty
> >>>>
> >>>> Migration region looks like:
> >>>>  ------------------------------------------------------------------
> >>>> |vfio_device_migration_info|    data section                      |
> >>>> |                          |     ///////////////////////////////  |
> >>>>  ------------------------------------------------------------------
> >>>>  ^                              ^                              ^
> >>>>  offset 0-trapped part        data_offset                 data_size
> >>>>
> >>>> Data section is always followed by vfio_device_migration_info
> >>>> structure in the region, so data_offset will always be none-0.
> >>>> Offset from where data is copied is decided by kernel driver, data
> >>>> section can be trapped or mapped depending on how kernel driver
> >>>> defines data section. If mmapped, then data_offset should be page
> >>>> aligned, where as initial section which contain
> >>>> vfio_device_migration_info structure might not end at offset which
> >>>> is page aligned.
> >>>>
> >>>> Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
> >>>> Reviewed-by: Neo Jia <cjia@nvidia.com>
> >>>> ---
> >>>>  linux-headers/linux/vfio.h | 71 ++++++++++++++++++++++++++++++++++++++++++++++
> >>>>  1 file changed, 71 insertions(+)
> >>>>
> >>>> diff --git a/linux-headers/linux/vfio.h b/linux-headers/linux/vfio.h
> >>>> index 24f505199f83..274ec477eb82 100644
> >>>> --- a/linux-headers/linux/vfio.h
> >>>> +++ b/linux-headers/linux/vfio.h
> >>>> @@ -372,6 +372,77 @@ struct vfio_region_gfx_edid {
> >>>>   */
> >>>>  #define VFIO_REGION_SUBTYPE_IBM_NVLINK2_ATSD	(1)
> >>>>  
> >>>> +/* Migration region type and sub-type */
> >>>> +#define VFIO_REGION_TYPE_MIGRATION	        (2)
> >>>> +#define VFIO_REGION_SUBTYPE_MIGRATION	        (1)
> >>>> +
> >>>> +/**
> >>>> + * Structure vfio_device_migration_info is placed at 0th offset of
> >>>> + * VFIO_REGION_SUBTYPE_MIGRATION region to get/set VFIO device related migration
> >>>> + * information. Field accesses from this structure are only supported at their
> >>>> + * native width and alignment, otherwise should return error.
> >>>> + *
> >>>> + * device_state: (read/write)
> >>>> + *      To indicate vendor driver the state VFIO device should be transitioned
> >>>> + *      to. If device state transition fails, write to this field return error.
> >>>> + *      It consists of 3 bits:
> >>>> + *      - If bit 0 set, indicates _RUNNING state. When its reset, that indicates
> >>>> + *        _STOPPED state. When device is changed to _STOPPED, driver should stop
> >>>> + *        device before write returns.
> >>>> + *      - If bit 1 set, indicates _SAVING state.
> >>>> + *      - If bit 2 set, indicates _RESUMING state.
> >>>> + *
> >>>> + * pending bytes: (read only)
> >>>> + *      Read pending bytes yet to be migrated from vendor driver
> >>>> + *
> >>>> + * data_offset: (read only)
> >>>> + *      User application should read data_offset in migration region from where
> >>>> + *      user application should read data during _SAVING state or write data
> >>>> + *      during _RESUMING state.
> >>>> + *
> >>>> + * data_size: (read/write)
> >>>> + *      User application should read data_size to know data copied in migration
> >>>> + *      region during _SAVING state and write size of data copied in migration
> >>>> + *      region during _RESUMING state.
> >>>> + *
> >>>> + * start_pfn: (write only)
> >>>> + *      Start address pfn to get bitmap of dirty pages from vendor driver duing
> >>>> + *      _SAVING state.
> >>>> + *
> >>>> + * page_size: (write only)
> >>>> + *      User application should write the page_size of pfn.
> >>>> + *
> >>>> + * total_pfns: (write only)
> >>>> + *      Total pfn count from start_pfn for which dirty bitmap is requested.
> >>>> + *
> >>>> + * copied_pfns: (read only)
> >>>> + *      pfn count for which dirty bitmap is copied to migration region.
> >>>> + *      Vendor driver should copy the bitmap with bits set only for pages to be
> >>>> + *      marked dirty in migration region.
> >>>> + *      Vendor driver should return 0 if there are 0 pages dirty in requested
> >>>> + *      range.
> >>>> + *      Vendor driver should return -1 to mark all pages in the section as
> >>>> + *      dirty.    
> >>>
> >>> Is the protocol that the user writes start_pfn/page_size/total_pfns in
> >>> any order and then the read of copied_pfns is what triggers the
> >>> snapshot?    
> >>
> >> Yes.
> >>  
> >>>  Are start_pfn/page_size/total_pfns sticky such that a user
> >>> can write them once and get repeated refreshes of the dirty bitmap by
> >>> re-reading copied_pfns?    
> >>
> >> Yes and that bitmap should be for given range (from start_pfn till
> >> start_pfn + tolal_pfns).
> >> Re-reading of copied_pfns is to handle the case where it might be
> >> possible that vendor driver reserved area for bitmap < total bitmap size
> >> for range (start_pfn to start_pfn + tolal_pfns), then user will have to
> >> iterate till copied_pfns == total_pfns or till copied_pfns == 0 (that
> >> is, there are no pages dirty in rest of the range)  
> > 
> > So reading copied_pfns triggers the data range to be updated, but the
> > caller cannot assume it to be synchronous and uses total_pfns to poll
> > that the update is complete?  How does the vendor driver differentiate
> > the user polling for the previous update to finish versus requesting a
> > new update?
> >   
> 
> Write on start_pfn/page_size/total_pfns, then read on copied_pfns
> indicates new update, where as sequential read on copied_pfns indicates
> polling for previous update.

Hmm, this seems to contradict the answer to my question above where I
ask if the write fields are sticky so a user can trigger a refresh via
copied_pfns.  Does it really make sense that this is asynchronous?  Are
we going to need to specify polling intervals and completion eventfds?
data_size is synchronous, right?  Thanks,

Alex

> >>>  What's the advantage to returning -1 versus
> >>> returning copied_pfns == total_pfns?
> >>>     
> >>
> >> If all bits in bitmap are 1, then return -1, that is, all pages in the
> >> given range to be marked dirty.
> >>
> >> If all bits in bitmap are 0, then return 0, that is, no page to be
> >> marked dirty in given range or rest of the range.
> >>
> >> Otherwise vendor driver should return copied_pfns == total_pfn and
> >> provide bitmap for total_pfn, which means that bitmap copied for given
> >> range contains information for all pages where some bits are 0s and some
> >> are 1s.  
> > 
> > Given that the vendor driver can indicate zero dirty pfns and all dirty
> > pfns, I interpreted copied_pfns as a synchronous operation where the
> > return value could indicate the number of dirty pages within the
> > requested range.
> >   
> >>> If the user then wants to switch back to reading device migration
> >>> state, is it a read of data_size that switches the data area back to
> >>> making that address space available?     
> >>
> >> No, Its not just read(data_size), before that there is a
> >> read(data_offset). If Vendor driver wants to have different sub-regions
> >> for device data and dirty page bitmap, vendor driver should return
> >> corresponding offset on read(data_offset).  
> > 
> > The dynamic use of data_offset was not at all evident to me until I got
> > further into the QEMU series.  The usage model needs to be well
> > specified in the linux header.  I infer this behavior is such that the
> > vendor driver can effectively identity map portions of device memory
> > and the user will restore to the same offset.  I suppose this is a
> > valid approach but it seems specifically tuned to devices which allow
> > full direct mapping, whereas many devices have more device memory than
> > is directly map'able and state beyond simple device memory.  Does this
> > model unnecessarily burden such devices?  It is a nice feature that
> > they data range can contain both mmap'd sections and trapped sections
> > and by adjusting data_offset the vendor driver can select which is
> > currently being used, but we really need to formalize all these details.
> >   
> >>> In each case, is it the user's
> >>> responsibility to consume all the data provided before triggering the
> >>> next data area?> For example, if I ask for a range of dirty bitmap, the
> >>> vendor driver will provide that range and and clear it, such that the
> >>> pages are considered clean regardless of whether the user consumed the
> >>> data area.      
> >>
> >> Yes.
> >>  
> >>> Likewise if the user asks for data_size, that would be
> >>> deducted from pending_bytes regardless of the user reading the data
> >>> area.     
> >>
> >> User should read data before deducting data_size from pending_bytes.  
> > 
> > The user deducts data_size form pending_bytes?  pending_bytes is
> > read-only, how does this work?  
> 
> Pending_bytes is readonly from migration region. User should read device
> data while pending_bytes > 0. How User would decide to iterate or not?
> User will have to check if previously read pending_bytes - data_size is
> still > 0, if yes then iterate. Before iterating, its users
> responsibility to read data from data section.
> 
> >   
> >> From vendor driver point of view, data_size will be deducted from
> >> pending_bytes once data is copied to data region.  
> > 
> > If the data is entirely from an mmap'd range, how does the vendor
> > driver know when the data is copied?
> >   
> >>> Are there any read side-effects to pending_bytes?    
> >>
> >> No, its query to vendor driver about pending bytes yet to be
> >> migrated/read from vendor driver.
> >>  
> >>>  Are there
> >>> read side-effects to the data area on SAVING?    
> >>
> >> No.  
> > 
> > So the vendor driver must make an assumption somewhere in the usage
> > protocol that it's the user's responsibility, this needs to be
> > specified.
> >   
> 
> Ok.
> 
> >>>  Are there write
> >>> side-effects on RESUMING, or is it only the write of data_size that
> >>> triggers the buffer to be consumed?    
> >>
> >> Its write(data_size) triggers the buffer to be consumed, if region is
> >> mmaped, then data is already copied to region, if its trapped then
> >> following writes from data_offset is data to be consumed.
> >>  
> >>>  Is it the user's responsibility to
> >>> write only full "packets" on RESUMING?  For example if the SAVING side
> >>> provides data_size X, that full data_size X must be written to the
> >>> RESUMING side, the user cannot write half of it to the data area on the
> >>> RESUMING side, write data_size with X/2, write the second half, and
> >>> again write X/2.  IOW, the data_size "packet" is indivisible at the
> >>> point of resuming.
> >>>     
> >>
> >> If source and destination are compatible or of same driver version, then
> >> if user is reading data_size X at source/SAVING, destination should be
> >> able to consume data_size X at restoring/RESUMING. Then why should user
> >> write X/2 and iterate?  
> > 
> > Because users do things we don't expect ;)  Maybe they decide to chunk
> > the data into smaller packets over the network, but the receiving side
> > would rather write the packet immediately rather than queuing it.
> > OTOH, does it necessarily matter so long as data_size is written on
> > completion of a full "packet"?
> >   
> 
> Doesn't matter. As long as data is written in same order as it was read,
> size doesn't matter.
> 
> >>> What are the ordering requirements?  Must the user write data_size
> >>> packets in the same order that they're read, or is it the vendor
> >>> driver's responsibility to include sequence information and allow
> >>> restore in any order?
> >>>     
> >>
> >> For user, data is opaque. User should write data in the same order as he
> >> received.  
> > 
> > Let's make sure that's specified.
> >   
> 
> Ok.
> 
> Thanks,
> Kirti
> 
> >>>> + */
> >>>> +
> >>>> +struct vfio_device_migration_info {
> >>>> +        __u32 device_state;         /* VFIO device state */
> >>>> +#define VFIO_DEVICE_STATE_STOPPED   (0)    
> >>>
> >>> We need to be careful with how this is used if we want to leave the
> >>> possibility of using the remaining 29 bits of this register.  Maybe we
> >>> want to define VFIO_DEVICE_STATE_MASK and be sure that we only do
> >>> read-modify-write ops within the mask (ex. set_bit and clear_bit
> >>> helpers).    
> >>
> >> Makes sense, I'll do changes in next iteration.
> >>  
> >>>  Also, above we define STOPPED to indicate simply
> >>> not-RUNNING, but here it seems STOPPED means not-RUNNING, not-SAVING,
> >>> and not-RESUMING.
> >>>     
> >>
> >> That's correct.
> >>  
> >>>> +#define VFIO_DEVICE_STATE_RUNNING   (1 << 0)
> >>>> +#define VFIO_DEVICE_STATE_SAVING    (1 << 1)
> >>>> +#define VFIO_DEVICE_STATE_RESUMING  (1 << 2)
> >>>> +#define VFIO_DEVICE_STATE_INVALID   (VFIO_DEVICE_STATE_SAVING | \
> >>>> +                                     VFIO_DEVICE_STATE_RESUMING)
> >>>> +        __u32 reserved;
> >>>> +        __u64 pending_bytes;
> >>>> +        __u64 data_offset;    
> >>>
> >>> Placing the data more than 4GB into the region seems a bit absurd, so
> >>> this could probably be a __u32 and take the place of the reserved field.
> >>>     
> >>
> >> Is there a maximum limit on VFIO region size?
> >> There isn't any such limit, right? Vendor driver can define region of
> >> any size and then place data section anywhere in the region. I prefer to
> >> keep it __u64.  
> > 
> > We have a single file descriptor for all accesses to the device, which
> > gives us quite a bit of per device address space.  As I mention above,
> > it wasn't clear to me that data_offset is used dynamically until I got
> > further into the series, so it seemed strange to me that we'd choose
> > such a large offset, but given my new understanding I agree it requires
> > a __u64 currently.  Thanks,
> > 
> > Alex
> >   
> >>>> +        __u64 data_size;
> >>>> +        __u64 start_pfn;
> >>>> +        __u64 page_size;
> >>>> +        __u64 total_pfns;
> >>>> +        __s64 copied_pfns;    
> >>>
> >>> If this is signed so that we can get -1 then the user could
> >>> theoretically specify total_pfns that we can't represent in
> >>> copied_pfns.  Probably best to use unsigned and specify ~0 rather than
> >>> -1.
> >>>     
> >>
> >> Ok.
> >>  
> >>> Overall this looks like a good interface, but we need to more
> >>> thoroughly define the protocol with the data area and set expectations
> >>> we're placing on the user and vendor driver.  There should be no usage
> >>> assumptions, it should all be spelled out.  Thanks,
> >>>    
> >>
> >> Thanks for your feedback. I'll update comments above to be more specific.
> >>
> >> Thanks,
> >> Kirti
> >>  
> >>> Alex
> >>>     
> >>>> +} __attribute__((packed));
> >>>> +
> >>>>  /*
> >>>>   * The MSIX mappable capability informs that MSIX data of a BAR can be mmapped
> >>>>   * which allows direct access to non-MSIX registers which happened to be within    
> >>>     
> >