[Qemu-devel] Re: AIO ide patch

[Qemu-devel] Re: AIO ide patch

[Vladimir N. Oleynik]

I wrote:

> Hi Fabrice!
> 
> I have tried to make a variant with asynchronous input/output
> for ide. First worked patch attached.
> (this variant haven`t AIO for ATAPI, only ATA).
> May be not full right, but ...

Good news: with this second variant patch installation w2k is full true
and don`t make disk full!
 From second variant I moved set_irq() to thread.
But, I don`t know: this save or not, but all my test cleans.


--w
vodz

Re: [Qemu-devel] Re: AIO ide patch

[Vladimir N. Oleynik]

Hetz,

>> Good news: with this second variant patch installation w2k is full true
>> and don`t make disk full!
>>  From second variant I moved set_irq() to thread.
>> But, I don`t know: this save or not, but all my test cleans.

> And where is the patch? :)

I want see Fabrice a reaction. I don`t know of safe my idea.
I tested 3-th variant with AIO for ATAPI. My tests is works.


ftp://ftp.simtreas.ru/pub/my/qemu/qemu_aio.patch3.gz


--w
vodz

Re: [Qemu-devel] Re: AIO ide patch

[Hetz Ben Hamo]

> Good news: with this second variant patch installation w2k is full true
> and don`t make disk full!
>  From second variant I moved set_irq() to thread.
> But, I don`t know: this save or not, but all my test cleans.

And where is the patch? :)

Thanks,
Hetz

[Qemu-devel] Re: AIO ide patch

[Fabrice Bellard]

Vladimir N. Oleynik wrote:
> I wrote:
> 
>> Hi Fabrice!
>>
>> I have tried to make a variant with asynchronous input/output
>> for ide. First worked patch attached.
>> (this variant haven`t AIO for ATAPI, only ATA).
>> May be not full right, but ...
> 
> 
> Good news: with this second variant patch installation w2k is full true
> and don`t make disk full!
>  From second variant I moved set_irq() to thread.
> But, I don`t know: this save or not, but all my test cleans.

The disk dull problem is timing dependent. Your patches change the 
timings, but there is no proof they really fixed the bug. Of course it 
is possible that Win2K has a bug which makes the IDE driver timing 
dependent...

Fabrice.

[Qemu-devel] Re: AIO ide patch

[Vladimir N. Oleynik]

Fabrice,

 >> I have tried to make a variant with asynchronous input/output
 >> for ide. First worked patch attached.
 >> (this variant haven`t AIO for ATAPI, only ATA).
 >> May be not full right, but ...

 >> Good news: with this second variant patch installation w2k is full true
 >> and don`t make disk full!
 >> From second variant I moved set_irq() to thread.
 >> But, I don`t know: this save or not, but all my test cleans.

 > The disk dull problem is timing dependent.

Yes.

 > Your patches change
 > the timings, but there is no proof they really fixed the bug.

I can`t patch W2k ;-)

 > Of course it is possible that Win2K has a bug which makes
 > the IDE driver timing dependent...

My idea is not patch timing. My idea is async IO for make speed up
(guest CPU continued after host`s disk/network/sound make IO processing).

If set_pic_irq() is tread safe, I can generate more other async IO
(floppy, DMA, sound, network fd poll)...


--w
vodz

Re: [Qemu-devel] Re: AIO ide patch

[Hetz Ben Hamo]

> I want see Fabrice a reaction. I don`t know of safe my idea.
> I tested 3-th variant with AIO for ATAPI. My tests is works.
>
>
> ftp://ftp.simtreas.ru/pub/my/qemu/qemu_aio.patch3.gz

Well, safe  - it's not :(

It seems that with your patch, Windows XP upon booting complains about wrong 
HAL, and doesn't boot. Windows 98 complains about corrupt hard drive and 
doesn't boot either.

With your patch applied - do I need to re-install the guest OS?

Thanks,
Hetz

Re: [Qemu-devel] Re: AIO ide patch

[Vladimir N. Oleynik]

Hetz,

>>I want see Fabrice a reaction. I don`t know of safe my idea.
>>I tested 3-th variant with AIO for ATAPI. My tests is works.
>>
>>
>>ftp://ftp.simtreas.ru/pub/my/qemu/qemu_aio.patch3.gz

> Well, safe  - it's not :(

May be. Qemu have comment for demontsrate this ;-)

/* XXX: tb_flush is currently not thread safe */
....
     if (cpu_single_env)
         env = cpu_single_env; /* XXX: find a correct solution for multithread */

And many other...

Current cpu_single_env=global_env and cpu_loop() have save context by localy,
not to env. io_port make irq from signle_env, mail_loop() make timer irq from
global_env, network fd_read make irq from mail_loop() with single_env...
I confused this. The current design may be ready not thread safe, but have
working progress for this.

> It seems that with your patch, Windows XP upon booting complains about wrong 
> HAL, and doesn't boot. Windows 98 complains about corrupt hard drive and 
> doesn't boot either.

Heh. My WinXP installation with AIO_patch4v is ok.
4v have async call all set_irq with usage sched_yield().

--w
vodz