[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH 00/12] add per-domain and per-cpupool generic parameters


  • To: Juergen Gross <jgross@xxxxxxxx>, Jan Beulich <JBeulich@xxxxxxxx>
  • From: George Dunlap <george.dunlap@xxxxxxxxxx>
  • Date: Tue, 18 Sep 2018 14:57:02 +0100
  • Autocrypt: addr=george.dunlap@xxxxxxxxxx; prefer-encrypt=mutual; keydata= xsFNBFPqG+MBEACwPYTQpHepyshcufo0dVmqxDo917iWPslB8lauFxVf4WZtGvQSsKStHJSj 92Qkxp4CH2DwudI8qpVbnWCXsZxodDWac9c3PordLwz5/XL41LevEoM3NWRm5TNgJ3ckPA+J K5OfSK04QtmwSHFP3G/SXDJpGs+oDJgASta2AOl9vPV+t3xG6xyfa2NMGn9wmEvvVMD44Z7R W3RhZPn/NEZ5gaJhIUMgTChGwwWDOX0YPY19vcy5fT4bTIxvoZsLOkLSGoZb/jHIzkAAznug Q7PPeZJ1kXpbW9EHHaUHiCD9C87dMyty0N3TmWfp0VvBCaw32yFtM9jUgB7UVneoZUMUKeHA fgIXhJ7I7JFmw3J0PjGLxCLHf2Q5JOD8jeEXpdxugqF7B/fWYYmyIgwKutiGZeoPhl9c/7RE Bf6f9Qv4AtQoJwtLw6+5pDXsTD5q/GwhPjt7ohF7aQZTMMHhZuS52/izKhDzIufl6uiqUBge 0lqG+/ViLKwCkxHDREuSUTtfjRc9/AoAt2V2HOfgKORSCjFC1eI0+8UMxlfdq2z1AAchinU0 eSkRpX2An3CPEjgGFmu2Je4a/R/Kd6nGU8AFaE8ta0oq5BSFDRYdcKchw4TSxetkG6iUtqOO ZFS7VAdF00eqFJNQpi6IUQryhnrOByw+zSobqlOPUO7XC5fjnwARAQABzSRHZW9yZ2UgVy4g RHVubGFwIDxkdW5sYXBnQHVtaWNoLmVkdT7CwYAEEwEKACoCGwMFCwkIBwMFFQoJCAsFFgID AQACHgECF4ACGQEFAlpk2IEFCQo9I54ACgkQpjY8MQWQtG1A1BAAnc0oX3+M/jyv4j/ESJTO U2JhuWUWV6NFuzU10pUmMqpgQtiVEVU2QbCvTcZS1U/S6bqAUoiWQreDMSSgGH3a3BmRNi8n HKtarJqyK81aERM2HrjYkC1ZlRYG+jS8oWzzQrCQiTwn3eFLJrHjqowTbwahoiMw/nJ+OrZO /VXLfNeaxA5GF6emwgbpshwaUtESQ/MC5hFAFmUBZKAxp9CXG2ZhTP6ROV4fwhpnHaz8z+BT NQz8YwA4gkmFJbDUA9I0Cm9D/EZscrCGMeaVvcyldbMhWS+aH8nbqv6brhgbJEQS22eKCZDD J/ng5ea25QnS0fqu3bMrH39tDqeh7rVnt8Yu/YgOwc3XmgzmAhIDyzSinYEWJ1FkOVpIbGl9 uR6seRsfJmUK84KCScjkBhMKTOixWgNEQ/zTcLUsfTh6KQdLTn083Q5aFxWOIal2hiy9UyqR VQydowXy4Xx58rqvZjuYzdGDdAUlZ+D2O3Jp28ez5SikA/ZaaoGI9S1VWvQsQdzNfD2D+xfL qfd9yv7gko9eTJzv5zFr2MedtRb/nCrMTnvLkwNX4abB5+19JGneeRU4jy7yDYAhUXcI/waS /hHioT9MOjMh+DoLCgeZJYaOcgQdORY/IclLiLq4yFnG+4Ocft8igp79dbYYHkAkmC9te/2x Kq9nEd0Hg288EO/OwE0EVFq6vQEIAO2idItaUEplEemV2Q9mBA8YmtgckdLmaE0uzdDWL9To 1PL+qdNe7tBXKOfkKI7v32fe0nB4aecRlQJOZMWQRQ0+KLyXdJyHkq9221sHzcxsdcGs7X3c 17ep9zASq+wIYqAdZvr7pN9a3nVHZ4W7bzezuNDAvn4EpOf/o0RsWNyDlT6KECs1DuzOdRqD oOMJfYmtx9hMzqBoTdr6U20/KgnC/dmWWcJAUZXaAFp+3NYRCkk7k939VaUpoY519CeLrymd Vdke66KCiWBQXMkgtMGvGk5gLQLy4H3KXvpXoDrYKgysy7jeOccxI8owoiOdtbfM8TTDyWPR Ygjzb9LApA8AEQEAAcLBZQQYAQoADwIbDAUCWmTXMwUJB+tP9gAKCRCmNjwxBZC0bb+2D/9h jn1k5WcRHlu19WGuH6q0Kgm1LRT7PnnSz904igHNElMB5a7wRjw5kdNwU3sRm2nnmHeOJH8k Yj2Hn1QgX5SqQsysWTHWOEseGeoXydx9zZZkt3oQJM+9NV1VjK0bOXwqhiQyEUWz5/9l467F S/k4FJ5CHNRumvhLa0l2HEEu5pxq463HQZHDt4YE/9Y74eXOnYCB4nrYxQD/GSXEZvWryEWr eDoaFqzq1TKtzHhFgQG7yFUEepxLRUUtYsEpT6Rks2l4LCqG3hVD0URFIiTyuxJx3VC2Ta4L H3hxQtiaIpuXqq2D4z63h6vCx2wxfZc/WRHGbr4NAlB81l35Q/UHyMocVuYLj0llF0rwU4Aj iKZ5qWNSEdvEpL43fTvZYxQhDCjQTKbb38omu5P4kOf1HT7s+kmQKRtiLBlqHzK17D4K/180 ADw7a3gnmr5RumcZP3NGSSZA6jP5vNqQpNu4gqrPFWNQKQcW8HBiYFgq6SoLQQWbRxJDHvTR YJ2ms7oCe870gh4D1wFFqTLeyXiVqjddENGNaP8ZlCDw6EU82N8Bn5LXKjR1GWo2UK3CjrkH pTt3YYZvrhS2MO2EYEcWjyu6LALF/lS6z6LKeQZ+t9AdQUcILlrx9IxqXv6GvAoBLJY1jjGB q+/kRPrWXpoaQn7FXWGfMqU+NkY9enyrlw==
  • Cc: Tim Deegan <tim@xxxxxxx>, Stefano Stabellini <sstabellini@xxxxxxxxxx>, Wei Liu <wei.liu2@xxxxxxxxxx>, Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>, George Dunlap <George.Dunlap@xxxxxxxxxxxxx>, Andrew Cooper <andrew.cooper3@xxxxxxxxxx>, Ian Jackson <Ian.Jackson@xxxxxxxxxxxxx>, Dario Faggioli <dfaggioli@xxxxxxxx>, Julien Grall <julien.grall@xxxxxxx>, xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxxx>, Daniel de Graaf <dgdegra@xxxxxxxxxxxxx>
  • Delivery-date: Tue, 18 Sep 2018 13:57:12 +0000
  • List-id: Xen developer discussion <xen-devel.lists.xenproject.org>
  • Openpgp: preference=signencrypt

On 09/18/2018 02:36 PM, Juergen Gross wrote:
> On 18/09/18 15:25, George Dunlap wrote:
>> On 09/18/2018 12:32 PM, Juergen Gross wrote:
>>> On 18/09/18 13:20, Jan Beulich wrote:
>>>>>>> On 18.09.18 at 13:10, <jgross@xxxxxxxx> wrote:
>>>>> On 18/09/18 12:32, Jan Beulich wrote:
>>>>>>>>> On 18.09.18 at 08:02, <jgross@xxxxxxxx> wrote:
>>>>>>> Instead of using binary hypervisor interfaces for new parameters of
>>>>>>> domains or cpupools this patch series adds support for generic text
>>>>>>> based parameter parsing.
>>>>>>>
>>>>>>> Parameters are defined via new macros similar to those of boot
>>>>>>> parameters. Parsing of parameter strings is done via the already
>>>>>>> existing boot parameter parsing function which is extended a little
>>>>>>> bit.
>>>>>>>
>>>>>>> Parameter settings can either be specified in configuration files of
>>>>>>> domains or cpupools, or they can be set via new xl sub-commands.
>>>>>>
>>>>>> Without having looked at any of the patches yet (not even their
>>>>>> descriptions) I'm still wondering what the benefit of textual parameters
>>>>>> really is: Just like "binary" ones, they become part of the public
>>>>>> interface, and hence subsequently can't be changed any more or
>>>>>> less than the ones we currently have (in particular, anything valid in
>>>>>> a guest config file will imo need to remain to be valid and meaningful
>>>>>> down the road).
>>>>>
>>>>> So lets look what would be needed for adding something like the
>>>>> per-domain xpti parameter using binary interfaces:
>>>>>
>>>>> 1 an extension of some domctl interface, maybe bumping of the domctl
>>>>>   interface version
>>>>> 2 adding the logic to domctl handling
>>>>> 3 adding libxc support
>>>>> 4 adding libxl support
>>>>> 5 adding a new xl sub-command
>>>>> 6 adding domain config support
>>>>> 7 adding documentation
>>>>>
>>>>> With my approach only 2 (in a modified form, parameter handling instead
>>>>> of domctl, but comparable in the needed effort) and 7 are needed.
>>>>>
>>>>> So once the framework is in place it is _much_ easier to add new
>>>>> features.
>>>>
>>>> All the above would hold if the individual options were expressed as
>>>> e.g. flags in an extensible bit vector.
>>>
>>> Who would translate the new option into a bit vector? This would be the
>>> tools (libxc/libxl/xl), so those need to be modified for each new bit.
>>
>> A bit vector would only allow on/off; it wouldn't allow you to set
>> numeric parameters, for example.
>>
>> I like the idea of being able to add configuration parameters without
>> having a huge amount of boilerplate; and also of being able to backport
>> parameters like xpti without having to worry so much about compatibility.
>>
>> But I'm not a fan of the idea of using a "string blob" to accomplish
>> that.  It's convenient for the exact use case you seem to be
>> contemplating: having a user add the string into the xl config file, and
>> having nobody but the hypervisor interpret it.  But what about tools
>> that may want to set that parameter?  Or tools that want to query the
>> parameter, or "introspect" on the domain settings or whatever?  Now they
>> have to have a bunch of code to generate and parse the string code.
>>
>> Could we have a reasonably generic structure / union, with a parameter
>> number, that we could pass in instead?  Something like:
>>
>> struct domain_parameter {
>>   int param_num;
>>   int val;
>> }
>>
>> And then have a list somewhere of string values -> parameter numbers
>> that callers could use to translate strings into values?
>>
>> That way the above list would look like:
>>
>> 1. Add new parameter in Xen
>> 2. Add a string name -> parameter number in a header somewhere
>> 3. Add a libxl #define with that parameter number
>>
>> You'd have to recompile xl against the new header, but you were probably
>> going to do that anyway.
> 
> The string variant is much more flexible.
> 
> It is easy possible to e.g. add a per-domain trace parameter to specify
> rather complex trace instrumentations. Doing something like that via a
> struct based interface is in the best case complicated.

...or, for instance, specifying the runqueue layout of a cpupool (for
schedulers like credit2 which allow such things).  Yes, that is true;
but probably a very niche case.

> Another advantage of the string based variant is that you don't need a
> central header. You can define the parameters where you are implementing
> them. No need to expand switch statements and headers, just a local
> definition and maybe a handler function.

I don't see the lack of central header as a big advantage -- how hard is
it to add a single line to a list somewhere?

*Not* having a language-level construct around (either an enum or a
#define) means that programs can't take advantage of the preprocessor /
type system to catch bugs; someone calling
libxl_domain_param("xptl=off"); won't get a compile-time error, only a
run-time one; someone calling
libxl_domain_param(LIBXL_DOMAIN_PARAM_XPTL, 0) will.

I'm not completely opposed to the "string blob" idea, but it would be
nice if at least for the common case of simple boolean / integer values,
we could avoid having a string blob.

What about

struct parameter {
  int param_number;
  union {
    int val;
    char special[]
  };
}

Or something like that?  That would give flexibility for special cases
like mentioned above, while allowing the common case to avoid special
encoding / decoding &c.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.