Re: Block protocol incompatibilities with 4K logical sector size disks

[Anthony PERARD <anthony.perard@vates.tech>]

On Thu, Aug 29, 2024 at 12:59:43PM +0200, Roger Pau Monné wrote:
> Hello,
>
> To give some context, this started as a bug report against FreeBSD failing to
> work with PV blkif attached disks with 4K logical sectors when the backend is
> Linux kernel blkback:
>
> https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=280884
>
> Further investigation has lead me to discover that the protocol described in
> the public blkif.h header is not implemented uniformly, and there are major
> inconsistencies between implementations regarding the meaning of the `sectors`
> and `sector-size` xenstore nodes, and the sector_number and {first,last}_sect
> struct request fields.  Below is a summary of the findings on the
> implementation I've analyzed.
>
> Linux blk{front,back} always assumes the `sectors` xenstore node to be in 512b
> units, regardless of the value of the `sector-size` node.  Equally the ring
> request sector_number and the segments {first,last}_sect fields are always
> assumed to be in units of 512b regardless of the value of `sector-size`.  The
> `feature-large-sector-size` node is neither exposed by blkfront, neither
> checked by blkback before exposing a `sector-size` node different than 512b.
>
> FreeBSD blk{front,back} calculates (and for blkback exposes) the disk size as
> `sectors` * `sector-size` based on the values in the xenstore nodes (as
> described in blkif.h).  The ring sector_number is filled with the sector number
> based on the `sector-size` value, however the {first,last}_sect fields are
> always calculated as 512b units.   The `feature-large-sector-size` node is
> neither exposed by blkfront, neither checked by blkback before exposing a
> `sector-size` node different than 512b.
>
> QEMU qdisk blkback implementation exposes the `sectors` disk size in units of
> `sector-size` (as FreeBSD blkback).  The ring structure fields sector_number
> and {first,last}_sect are assumed to be in units of `sector-size`.  This
> implementation will not expose a `sector-size` node with a value different than
> 512 unless the frontend xenstore path has the `feature-large-sector-size` node
> present.
>
> Windows blkfront calculates the disk size as `sectors` * `sector-size` from the
> xenstore nodes exposed by blkback.   The ring structure fields sector_number
> and {first,last}_sect are assumed to be in units of `sector-size`.  This
> frontend implementation exposes `feature-large-sector-size`.
>
> When using a disk with a logical sector size different than 512b, Linux is only
> compatible with itself, same for FreeBSD.  QEMU blkback implementation is also
> only compatible with the Windows blkfront implementation.  The
> `feature-large-sector-size` seems to only be implemented for the QEMU/Windows
> combination, both Linux and FreeBSD don't implement any support for it neither
> in the backend or the frontend.
>
> The following table attempts to summarize in which units the following fields
> are defined for the analyzed implementations (please correct me if I got some
> of this wrong):
>
>                         │ sectors xenbus node │ requests sector_number │ requests {first,last}_sect
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> FreeBSD blk{front,back} │     sector-size     │      sector-size       │           512
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> Linux blk{front,back}   │         512         │          512           │           512
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> QEMU blkback            │     sector-size     │      sector-size       │       sector-size
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> Windows blkfront        │     sector-size     │      sector-size       │       sector-size
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> MiniOS                  │     sector-size     │          512           │           512
> ────────────────────────┼─────────────────────┼────────────────────────┼───────────────────────────
> tapdisk blkback         │         512         │      sector-size       │           512

There's OVMF as well, which copied MiniOS's implementation, and looks
like it's still the same as MiniOS for the table above:

OVMF (base on MiniOS)   │     sector-size     │          512           │           512

>
> It's all a mess, I'm surprised we didn't get more reports about brokenness when
> using disks with 4K logical sectors.
>
> Overall I think the in-kernel backends are more difficult to update (as it
> might require a kernel rebuild), compared to QEMU or blktap.  Hence my slight
> preference would be to adjust the public interface to match the behavior of
> Linux blkback, and then adjust the implementation in the rest of the backends
> and frontends.

I would add that making "sector-size" been different from 512 illegal
makes going forward easier, has every implementation will work with a
"sector-size" of 512, and it probably going to be the most common sector
size for a while longer.

> There was an attempt in 2019 to introduce a new frontend feature flag to signal
> whether the frontend supported `sector-size` xenstore nodes different than 512 [0].
> However that was only ever implemented for QEMU blkback and Windows blkfront,
> all the other backends will expose `sector-size` different than 512 without
> checking if `feature-large-sector-size` is exposed by the frontend.  I'm afraid
> it's now too late to retrofit that feature into existing backends, seeing as
> they already expose `sector-size` nodes greater than 512 without checking if
> `feature-large-sector-size` is reported by the frontend.

Much before that, "physical-sector-size" was introduced (2013):
    https://xenbits.xen.org/gitweb/?p=xen.git;a=commit;h=a67e2dac9f8339681b30b0f89274a64e691ea139

Linux seems to implement it, but QEMU or OVMF don't have it.

> My proposal would be to adjust the public interface with:
>
>  * Disk size is calculated as: `sectors` * 512 (`sectors` being the contents of
>    such xenstore backend node).
>
>  * All the sector related fields in blkif ring requests use a 512b base sector
>    size, regardless of the value in the `sector-size` xenstore node.
>
>  * The `sector-size` contains the disk logical sector size.  The frontend must
>    ensure that all request segments addresses are aligned and it's length is
>    a multiple of such size.  Otherwise the backend will refuse to process the
>    request.

You still want to try to have a "sector-size" different from 512? To me
this just add confusion to the confusion. There would be no way fro
backend or frontend to know if setting something other than 512 is going
to work. Also, it is probably easier to update backend than frontend, so
it is just likely that something is going to lag behind and broke.

Why not make use of the node "physical-sector-size" that have existed
for 10 years, even if unused or unadvertised, and if an IO request isn't
aligned on it, it is just going to be slow (as backend would have to
read,update,write instead of just write sectors).

Cheers,

--

Anthony Perard | Vates XCP-ng Developer

XCP-ng & Xen Orchestra - Vates solutions

web: https://vates.tech