Re: raw disks no longer work in latest kvm (kvm-88 was fine)

[Avi Kivity <avi@redhat.com>]

On 03/07/2010 08:01 PM, Antoine Martin wrote:
> On 03/08/2010 12:30 AM, Avi Kivity wrote:
>> On 03/07/2010 07:21 PM, Christoph Hellwig wrote:
>>> On Sun, Mar 07, 2010 at 07:18:40PM +0200, Avi Kivity wrote:
>>>>> The only things that stands out is this before the "read failed" 
>>>>> message:
>>>>> [pid  9098] lseek(12, 0, SEEK_END)      = 1321851815424
>>>>> [pid  9121] pread(12, 0x7fa50a0e47d0, 2048, 0) = -1 EINVAL (Invalid
>>>>> argument)
>>>>>
>>>> The buffer is unaligned here, yet the file was opened with O_DIRECT
>>>> (cache=none).  This is strange, since alignment is not related to disk
>>>> size.
> Yes cache=none: "-drive file=./vm/var_fs,if=virtio,cache=none"
>
> Side question: this is the right thing to do for raw partitions, right?

The rightest.

>>> The other interesting thing is that it's using pread - which means
>>> it's a too old kernel to use preadv and thus a not very well tested
>>> codepath with current qemu.
> Too old?, I am confused: both host and guest kernels are 2.6.33!
> I built KVM against the 2.6.30 headers though.

You need to build qemu against the 2.6.33 headers ('make 
headers-install').  But after we fix this, please.

>>
>> It may also be that glibc is emulating preadv, incorrectly.
> Not sure how to do that.
>>
>> Antoine, can you check this?  ltrace may help,
> This the only part that looked slightly relevant:
> [pid 26883] __xstat64(1, "./vm/media_fs", 0x7fff92e40500) = 0
> [pid 26883] malloc(48)                           = 0x2a38d60
> [pid 26883] memset(0x2a38d60, '\000', 48)        = 0x2a38d60
> [pid 26883] open64("./vm/media_fs", 540674, 00)  = 12
> [pid 26883] posix_memalign(0x7fff92e40560, 512, 16384, -1, 48) = 0
> [pid 26883] lseek64(12, 0, 2, 0x7f404f2f3e60, 4) = 0x133c4820600


Where's pread/preadv?  Did you use -f?

>
>> or run 'strings libc.so | grep pread'.
>>
> strings /lib/libc.so.6 | grep pread
> preadv
> preadv64
> pread
> __pread64_chk
> __pread64
> __pread_chk
>

So it does seem glibc emulates preadv.

Perhaps https://bugzilla.redhat.com/show_bug.cgi?id=563103 ?

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.