Re: [Qemu-devel] [PATCH v3 1/5] VFIO KABI for migration interface

[Alex Williamson]

On Wed, 20 Feb 2019 02:53:16 +0530
Kirti Wankhede <address@hidden> wrote:

> - Defined MIGRATION region type and sub-type.
> - Used 2 bits to define VFIO device states.
>     Bit 0 => 0/1 => _STOPPED/_RUNNING
>     Bit 1 => 0/1 => _RESUMING/_SAVING
>     Combination of these bits defines VFIO device's state during migration
>     _RUNNING => Normal VFIO device running state.
>     _STOPPED => VFIO device stopped.
>     _SAVING | _RUNNING => vCPUs are running, VFIO device is running but start
>                           saving state of device i.e. pre-copy state
>     _SAVING | _STOPPED => vCPUs are stoppped, VFIO device should be stopped, 
> and
>                           save device state,i.e. stop-n-copy state
>     _RESUMING => VFIO device resuming state.

Shouldn't we have a non-_RESUMING/_SAVING run state?  If these are
indicating directly flow, maybe we need two bits:

  00b - None, normal runtime
  01b - Saving
  10b - Resuming
  11b - Invalid/reserved (maybe a Failed state indicator)

> - 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: (write only)
>         To convey VFIO device state to be transitioned to.

Seems trivial and potentially useful to support read here, we have 30
(or maybe 29) bits yet to define.

>     * pending bytes: (read only)
>         To get pending bytes yet to be migrated for VFIO device
>     * data_offset: (read/write)
>         To get or set data offset in migration from where data exist
>         during _SAVING and _RESUMING state

What's the use case for writing this?

>     * data_size: (write only)
>         To convey size of data copied in migration region during _RESUMING
>         state

How to know how much is available for read?

>     * start_pfn, page_size, total_pfns: (write only)
>         To get bitmap of dirty pages from vendor driver from given
>         start address for total_pfns.

What would happen if a user wrote in 1MB for page size?  Is the vendor
driver expected to support arbitrary page sizes?  Are we only trying to
convey the page size and would that page size ever be other than
getpagesize()?

>     * 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.

This is useful, but I wonder if it's really a required feature for the
vendor driver.  For instance, with mdev IOMMU support we could wrap an
arbitrary PCI device as mdev, but we don't necessarily have dirty page
tracking.  Would a device need to report -1 here if it wanted to
indicate any page could be dirty if we only know how to collect the
state of the device itself for migration (ie. force the device to be
stopped first).
 
> Migration region looks like:
>  ------------------------------------------------------------------
> |vfio_device_migration_info|    data section                      |
> |                          |     ///////////////////////////////  |
>  ------------------------------------------------------------------
                             ^ what's this?

>  ^                              ^                              ^
>  offset 0-trapped part        data.offset                 data.size

Isn't data.size above really (data.offset + data.size)?  '.' vs '_'
inconsistency vs above.
 
> Data section is always followed by vfio_device_migration_info
> structure in the region, so data.offset will always be none-0.

This seems exactly backwards from the diagram, data section follows
vfio_device_migration_info.  Also, "non-zero".

> Offset from where data is copied is decided by kernel driver, data

But data_offset is listed as read-write.

> 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 <address@hidden>
> Reviewed-by: Neo Jia <address@hidden>
> ---
>  linux-headers/linux/vfio.h | 65 
> ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 65 insertions(+)
> 
> diff --git a/linux-headers/linux/vfio.h b/linux-headers/linux/vfio.h
> index 12a7b1dc53c8..1b12a9b95e00 100644
> --- a/linux-headers/linux/vfio.h
> +++ b/linux-headers/linux/vfio.h
> @@ -368,6 +368,71 @@ 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: (write only)
> + *      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 2 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. When its reset, that 
> indicates
> + *        _RESUMING state.
> + *
> + * pending bytes: (read only)
> + *      Read pending bytes yet to be migrated from vendor driver

Is this essentially a volatile value, changing based on data previously
copied and device activity?

> + *
> + * data_offset: (read/write)
> + *      User application should read data_offset in migration region from 
> where
> + *      user application should read data during _SAVING state.
> + *      User application would write data_offset in migration region from 
> where
> + *      user application is had written data during _RESUMING state.

Why wouldn't data_offset simply be fixed?  Do we really want to support
the user writing a arbitrary offsets in the migration region?  Each
chunk can simply start at the kernel provided data_offset.

> + *
> + * data_size: (write only)
> + *      User application should write size of data copied in migration region
> + *      during _RESUMING state.

How much does the user read on _SAVING then?

> + *
> + * 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.

So effectively the area that begins at data_offset is dual purpose
(triple purpose vs Yan's, config, memory, and dirty bitmap) but the
protocol isn't entirely evident to me. I think we need to write it out
as Yan provided.  If I'm in the _SAVING state, do I simply read from
data_offset to min(data_size, region.size - data_offset)?  If that area
is mmap'd, how does the user indicate to the kernel to prepare the data
or that X bytes were acquired?  In the _RESUMING state, do I write from
data_offset to min(data_size, region.size - data_offset) and indicate
the write using data_size?

> + *
> + * 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.
> + */
> +
> +struct vfio_device_migration_info {
> +        __u32 device_state;         /* VFIO device state */
> +#define VFIO_DEVICE_STATE_RUNNING   (1 << 0)
> +#define VFIO_DEVICE_STATE_SAVING    (1 << 1)
> +        __u32 reserved;
> +        __u64 pending_bytes;
> +        __u64 data_offset;
> +        __u64 data_size;
> +        __u64 start_pfn;
> +        __u64 page_size;
> +        __u64 total_pfns;
> +        __u64 copied_pfns;
> +} __attribute__((packed));
> +

As you mentioned, and with Yan's version, we still need to figure out
something with compatibility and versioning.  Thanks,

Alex

>  /*
>   * 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