[Libguestfs] RFC: nbdkit block size advertisement

[Eric Blake]

On 7/23/20 5:00 PM, Richard W.M. Jones wrote:
> On Thu, Jul 23, 2020 at 10:54:31AM -0500, Eric Blake wrote:
>> I'm thinking of adding one or more callbacks to nbdkit to let
>> plugins/filters enforce various block size alignments (for example,
>> the swab filter requires 2/4/8 alignment, or VDDK requires 512
>> alignment, etc).  The NBD protocol currently has NBD_INFO_BLOCK_SIZE
>> which can be sent in reply to NBD_OPT_GO to tell the client about
>> sizing constraints; qemu already implements it as both client and
>> server, so we have some reasonable testing setups (although libnbd
>> will also need some additions to make it easier to expose
>> constraints to the user, and/or add new convenience APIs to do
>> blocksize-style read-modify-write at the libnbd client side rather
>> than needing the blocksize filter in the NBD server side).
>>
>> But NBD_INFO_BLOCK_SIZE is not the full picture: it only covers
>> minimum block size, preferred block size, and maximum data size;
>> there has been discussion on the NBD list about also advertising
>> maximum action size (discussion has mentioned trim and/or zero, but
>> even cache could benefit from larger buffer size than pread), which
>> means we should be thinking about supporting future protocol
>> extensions in whatever we expose to plugins.
>>
>> So, I'm thinking something like the following:
>>
>> New enum:
>> NBDKIT_BLOCK_SIZE_GET_MINIMUM
>> NBDKIT_BLOCK_SIZE_GET_PREFERRED
>> NBDKIT_BLOCK_SIZE_GET_MAX_DATA
>> NBDKIT_BLOCK_SIZE_GET_MAX_TRIM
>> NBDKIT_BLOCK_SIZE_GET_MAX_ZERO
>> NBDKIT_BLOCK_SIZE_GET_MAX_CACHE
> 
> enum or int?  I think there are ABI problems with enums, although
> probably not unless we have more than 256 cases?

Definitely typed as 'int' in the .block_size prototype (you're right 
that enums and ABI is not a game I want to play with); but whether we 
use enum or a series of #define to give values to those constants is a 
matter of taste.  Porting to other language bindings, where that 
language has saner enum support than C, can of course use an enum.

> 
>> along with a new callback for plugins:
>>
>> int64_t block_size (void *handle, int which);
>>
>> where 'which' is one of the enum values.  A future nbdkit might
>> request an enum value not recognized at the time the plugin was
>> compiled, so the recommended behavior is that a plugin returns -1 on
>> error, 0 to let nbdkit pick a sane default (including when 'which'
>> was unknown), or a positive value for an actual limit.  For now,
>> nbdkit would advertise MIN, PREFERRED, and MAX_DATA limits (to
>> clients that use NBD_OPT_GO), while the others are enforced only
>> internally.  The idea is that if the plugin installs a limit, a
>> client request that violates that limit will fail with EINVAL for
>> being unaligned, without even asking the plugin to try the response.
> 
> Isn't the plan that the server would try to resolve the problem -
> eg. by making a RMW request or splitting a request?  This would be
> especially the case for internal requests, but I could also see it
> having value for clients which either ignore the block size
> information, or don't implement it correctly/completely.

The blocksize filter can do that.  I'm leaning more towards having the 
core server ignore the issue with -EINVAL, then use the blocksize filter 
when you want the server to fix things on your behalf, rather than 
bloating the core server when most clients are already fairly 
well-behaved (qemu, for example, generally sticks to 512-byte 
boundaries).  Also, the blocksize filter currently has the limitation 
that it serializes all requests (I was lazy and wrote it with a single 
bounce buffer shared among all requests, rather than allocating a bounce 
buffer per unaligned I/O).  There's also locking questions to worry 
about: suppose you have a plugin that wants 4k alignment, but a client 
sends two non-overlapping 512-byte writes in parallel to the same 4k 
portion of the disk.  We _really_ want our rmw sequences to be atomic. 
That is,

      thread 1                thread 2
start handling write(len=512, offset=0)
                           start handling write(len=512, offset=512)
widen to .pread(len=4k, offset=0)
                           widen to .pread(len=4k, offset=0)
modify buffer
                           modify buffer
commit with .pwrite(len=4k, offset=0)
                           commit with .pwrite(len=4k, offset=0)

is a BAD example of not locking, because thread 2 undoes the work by 
thread 1, even though the client had no overlap in its parallel requests 
and thus no reason to expect that corruption.


do not want to be causing data corruption during our rmw cycle, so we 
have to stall any secondary I/O request that overlaps with a request 
where we are already using a bounce buffer.

> 
>> nbdkit calls the plugin callback once per connection per reasonable
>> 'which' (caching the results like it does for .get_size, and
>> skipping limits where .can_FOO fails). Returning int64_t allows us
>> to reuse this callback without change when we add v3 callbacks,
>> although values larger than 4G make no difference at present.  I
>> thought the use of an enum was nicer than filling in a struct whose
>> size might change, or adding one callback per limit.
> 
> Yes the enum/int seems like a better idea than dealing with structs,
> and is also easier to map into other programming languages.  The
> overhead of a few extra indirect function calls is negligible because
> it's only once per connection.
> 
> The only other alternative I can see would be to put these into struct
> nbdkit_plugin (as int64_t fields), but we have always regretted using
> plain fields instead of functions in this struct, eg. thread_model,
> config_help, etc.

Yeah, I can _totally_ see these values being data-dependent (for 
example, in the file plugin, based on whether you used an option for 
turning on O_DIRECT, and whether the underlying filesystem advertises a 
preferred size).  So a function is the only way to keep it 
runtime-adjustable.

> 
>> Filters can relax limits (such as blocksize turning a plugin MIN 512
>> into an advertised MIN 1, by doing read-modify-write as needed) or
>> tighten limits (the file plugin has MIN 1, but the swab filter
>> imposes a tighter MIN 2).  Constraints between limits are as
>> follows:
> 
> Can the server do this transparently between layers?  Might be a lot
> easier to implement it once.

Well, it's only filters that would ever tweak things, and even then, 
there aren't that many filters that want to alter alignments.  I guess I 
just have to code up a first round to see how it looks.

> 
>> MIN: must be power of 2, between 1 and 64k; .get_size and .extents
>> must return results aligned to MIN (as any unaligned tail or extent
>> transition is inaccessible using only aligned operations).  Defaults
>> to 1.
>>
>> PREFERRED: must be power of 2, between max(512, MIN) and 2M (this
>> upper limit is not specified by NBD spec, but matches qemu's
>> implementation of what it uses as the max qcow2 cluster size).
>> Defaults to max(4k, MIN).
>>
>> MAX_DATA: must be multiple of MIN and at least as large as
>> PREFERRED; values larger than 64M are okay but clamped by nbdkit's
>> own internal limits. Defaults to 64M.
>>
>> MAX_TRIM, MAX_ZERO, MAX_CACHE: must be multiple of MIN, should be at
>> least as big as MAX_DATA. Values 4G and larger are clamped by
>> nbdkit's own internal limits. Defaults to 4G-MIN for TRIM/ZERO, and
>> MAX_DATA for CACHE.
> 
> Sounds reasonable otherwise.
> 
> Rich.
> 

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org