[Top][All Lists]

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

Re: [Qemu-devel] [RFC] block: Removed coroutine ownership assumption

From: Kevin Wolf
Subject: Re: [Qemu-devel] [RFC] block: Removed coroutine ownership assumption
Date: Mon, 02 Jul 2012 12:52:36 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:13.0) Gecko/20120605 Thunderbird/13.0

Am 02.07.2012 12:18, schrieb Peter Crosthwaite:
> On Mon, Jul 2, 2012 at 8:01 PM, Kevin Wolf <address@hidden> wrote:
>> Am 02.07.2012 11:42, schrieb Peter Crosthwaite:
>>> On Mon, Jul 2, 2012 at 7:04 PM, Kevin Wolf <address@hidden> wrote:
>>>> Am 02.07.2012 10:57, schrieb Peter Crosthwaite:
>>>>> No conditional on the qemu_coroutine_create. So it will always create
>>>>> a new coroutine for its work which will solve my problem. All I need
>>>>> to do is pump events once at the end of machine model creation. Any my
>>>>> coroutines will never yield or get queued by block/AIO. Sound like a
>>>>> solution?
>>>> If you don't need the read data in your initialisation code,
>>> definately not :) Just as long as the read data is there by the time
>>> the machine goes live.  Whats the current policy with bdrv_read()ing
>>> from init functions anyway? Several devices in qemu have init
>>> functions that read the entire storage into a buffer (then the guest
>>> just talks to the buffer rather than the backing store).
>> Reading from block devices during device initialisation breaks
>> migration, so I'd like to see it go away wherever possible. Reading in
>> the whole image file doesn't sound like something for which a good
>> excuse exists,
> Makes sense for small devices on embedded platforms. You end up with a
> very simple and clean device model. The approach is totally broken for
> HDDs but it does make some sense for the little embedded flashes where
> you can get away with caching the entire device storage in RAM for the
> lifetime of the system.

It kind of works for read-only devices, it's just ugly there. With
writeable devices it makes the VM unmigratable.

>> you can do that as well during the first access.
> Kind of painful though to change this for a lot of existing devices.
> Its a reasonable policy for new devices, but can we just fix the
> init->bdrv_read() calls in place for the moment?

Sure, but you asked for the policy and the policy is "only with good
reasons". ;-)

>>> Pflash (pflash_cfi_01.c) is the device that is causing me interference
>>> here and it works exactly like this. If we make the bdrv_read() aio
>>> though, how do you ensure that it has completed before the guest talks
>>> to the device? Will this just happen at the end of machine_init
>>> anyways? Can we put a one liner in the machine init framework that
>>> pumps all AIO events then just mass covert all these bdrv_reads (in
>>> init functions) to bdrv_aio_read with a nop completion callback?
>> The initialisation function of the device can wait at its end for all
>> AIOs to return.
> You lose the performance increase discussed below however, as instead
> of the init function returning to continue on with the machine init,
> you block on disk IO.

Do you think it really matters? I guess it might if you have two such
devices that take each a second or so to load, but otherwise there isn't
much to parallelise.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]