Re: [Libguestfs] [PATCH 1/2] spec: Recommend cap on NBD_REPLY_TYPE_BLOCK_STATUS length

[Eric Blake]

On Fri, Apr 08, 2022 at 04:48:59PM +0300, Nir Soffer wrote:
> On Fri, Apr 8, 2022 at 2:52 PM Richard W.M. Jones <rjones@redhat.com> wrote:
> >
> > On Fri, Apr 08, 2022 at 09:25:01AM +0200, Wouter Verhelst wrote:
> > > Hi Eric,
> > >
> > > On Thu, Apr 07, 2022 at 04:37:19PM -0500, Eric Blake wrote:
> > > > The spec was silent on how many extents a server could reply with.
> > > > However, both qemu and nbdkit (the two server implementations known to
> > > > have implemented the NBD_CMD_BLOCK_STATUS extension) implement a hard
> > > > cap, and will truncate the amount of extents in a reply to avoid
> > > > sending a client a reply larger than the maximum NBD_CMD_READ response
> > > > they are willing to tolerate:
> > > >
> > > > When qemu first implemented NBD_CMD_BLOCK_STATUS for the
> > > > base:allocation context (qemu commit e7b1948d51, Mar 2018), it behaved
> > > > as if NBD_CMD_FLAG_REQ_ONE were always passed by the client, and never
> > > > responded with more than one extent.  Later, when adding its
> > > > qemu:dirty-bitmap:XXX context extension (qemu commit 3d068aff16, Jun
> > > > 2018), it added a cap to 128k extents (1M+4 bytes), and that cap was
> > > > applied to base:allocation once qemu started sending multiple extents
> > > > for that context as well (qemu commit fb7afc797e, Jul 2018).  Qemu
> > > > extents are never smaller than 512 bytes (other than an exception at
> > > > the end of a file whose size is not aligned to 512), but even so, a
> > > > request for just under 4G of block status could produce 8M extents,
> > > > resulting in a reply of 64M if it were not capped smaller.

Qemu as server imposed the limit because sending a 64M reply to a
block status request for 4G (where every 512 bytes alternates status)
would kill qemu as client (which refuses to accept responses from the
server larger than 32M).  Nothing in here about considerations for the
server.

> > > >
> > > > When nbdkit first implemented NBD_CMD_BLOCK_STATUS (nbdkit 4ca66f70a5,
> > > > Mar 2019), it did not impose any restriction on the number of extents
> > > > in the reply chunk.  But because it allows extents as small as one
> > > > byte, it is easy to write a server that can amplify a client's request
> > > > of status over 1M of the image into a reply over 8M in size, and it
> > > > was very easy to demonstrate that a hard cap was needed to avoid
> > > > crashing clients or otherwise killing the connection (a bad server
> > > > impacting the client negatively);

Likewise, THIS part of the commit message is about nbdkit limiting things to 
avoid killing the client.

> > > > unique to nbdkit's situation is the
> > > > fact that because it is designed for plugin server implementations,
> > > > not capping the length of extent also posed a problem to nbdkit as the
> > > > server (a client requesting large block status could cause the server
> > > > to run out of memory depending on the plugin providing the server
> > > > callbacks).  So nbdkit enforced a bound of 1M extents (8M+4 bytes,
> > > > nbdkit commit 6e0dc839ea, Jun 2019).

While this part is an incidental anecdote (we also had problems with
clients being able to kill the server, when the server was poorly
written to consume too much memory by heavily alternating status), but
does NOT change the content of the text change in the patch.

> > > >
> > > > Since the limit chosen by these two implementations is different, and
> > > > since nbdkit has versions that were not limited, add this as a SHOULD
> > > > NOT instead of MUST NOT constraint on servers implementing block
> > > > status.  It does not matter that qemu picked a smaller limit that it
> > > > truncates to, since we have already documented that the server may
> > > > truncate for other reasons (such as it being inefficient to collect
> > > > that many extents in the first place).  But documenting the limit now
> > > > becomes even more important in the face of a future addition of 64-bit
> > > > requests, where a client's request is no longer bounded to 4G and
> > > > could thereby produce even more than 8M extents for the corner case
> > > > when every 512 bytes is a new extent, if it were not for this
> > > > recommendation.
> > >
> > > It feels backwards to me to make this a restriction on the server side.
> > > You're saying there are server implementations that will be inefficient
> > > if there are more than 2^20 extents, and therefore no server should send
> > > more than those, even if it can do so efficiently.

The limitation is on the server to avoid killing clients.  Whether or
not the server happens to be inefficient if it could otherwise
generate more than 2^20 extents (as nbdkit was temporarily able to do)
is not mentioned in the spec change in the patch itself, but only in
the commit message.  I can do a better job of marking the server
effect as an interesting side effect.  Servers can ALWAYS choose to
truncate early to avoid exhausting their own resources; the important
part of this patch is that servers MUST truncate early to avoid
killing the client, even if the server has the resources to provide a
larger response up to the bounds of the client's request.

> > >
> > > Isn't it up to the server implementation to decide what can be done
> > > efficiently?
> > >
> > > Perhaps we can make the language about possibly reducing length of
> > > extens a bit stronger; but I don't think adding explicit limits for a
> > > server's own protection is necessary.
> >
> > I agree, but for a different reason.
> >
> > I think Eric should add language that servers can consider limiting
> > response sizes in order to prevent possible amplification issues
> > and/or simply overwhelming the client with work (bad server DoS
> > attacks against clients are a thing!), but I don't think it's
> > necessarily a "SHOULD" issue.

On the contrary, allowing a server to send anything larger than 32M in
a reply chunk is GOING to kill clients (since many existing clients
treat server packets larger than 32M as a denial of service attack).
But the client has no control over how much the server will choose to
reply with, other than the client can limit its initial size hint to
something small enough that even if the server causes an 8-fold
increase in size of the reply by alternating status for every byte,
the server cannot send more than the number of extents requested by
the client.  But forcing the client to limit their extent hints is
counterproductive, since MOST images don't have that many alternating
extents, and it causes more traffic overhead to break extent queries
into smaller chunks just to worry about corner cases.

> >
> > BTW attached is an nbdkit plugin that creates an NBD server that
> > responds with massive numbers of byte-granularity extents, in case
> > anyone wants to test how nbdkit and/or clients respond:
> >
> > $ chmod +x /var/tmp/lots-of-extents.py
> > $ /var/tmp/lots-of-extents.py -f
> >
> > $ nbdinfo --map nbd://localhost | head
> >          0           1    3  hole,zero
> >          1           1    0  data
> >          2           1    3  hole,zero
> >          3           1    0  data
> >          4           1    3  hole,zero
> >          5           1    0  data
> >          6           1    3  hole,zero
> >          7           1    0  data
> >          8           1    3  hole,zero
> >          9           1    0  data
> > $ nbdinfo --map --totals nbd://localhost
> >     524288  50.0%   0 data
> >     524288  50.0%   3 hole,zero
> 
> This is a malicious server. A good client will drop the connection when
> receiving the first 1 byte chunk.

Depends on the server.  Most servers don't serve 1-byte extents, and
the NBD spec even recommends that extents be at least 512 bytes in
size, and requires that extents be a multiple of any minimum block
size if one was advertised by the server.

But even though most servers don't have 1-byte extents does not mean
that the NBD protocol must forbid them.


> 
> The real issue here is not enforcing or suggesting a limit on the number of
> extent the server returns, but enforcing a limit on the minimum size of
> a chunk.
> 
> Since this is the network *block device* protocol it should not allow chunks
> smaller than the device block size, so anything smaller than 512 bytes
> should be invalid response from the server.

No, not an invalid response, but merely a discouraged one - and that
text is already present in the wording of NBD_CMD_BLOCK_STATUS.

> 
> Even the last chunk should not be smaller than 512 bytes. The fact that you
> can serve a file with size that is not aligned to 512 bytes does not mean that
> the export size can be unaligned to the logical block size. There are no real
> block devices that have such alignment so the protocol should not allow this.
> A good server will round the file size down the logical block size to avoid 
> this
> issue.
> 
> How about letting the client set a minimum size of a chunk? This way we
> avoid the issue of limiting the number of chunks. Merging small chunks
> is best done on the server side instead of wasting bandwidth and doing
> this on the client side.

The client can't set the minimum block size, but the server can
certainly advertise one, and must obey that advertisement.  Or are you
asking for a new extension where the client mandates what the minimum
granularity must be from the server in responses to NBD_CMD_READ and
NBD_CMD_BLOCK_STATUS, when the client wants a larger granularity than
what the server advertises?  That's a different extension than this
patch, but may be worth considering.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org