Re: [PATCH] libxl: stat the path for all non-qdisk backends (including unknown)

[George Dunlap <george.dunlap@eu.citrix.com>]

On 10/05/13 16:03, Dave Scott wrote:
> On 10/05/13 15:42, George Dunlap wrote:
>> On 10/05/13 15:31, Dave Scott wrote:
>>> On 10/05/13 15:19, George Dunlap wrote:
>>>> On 10/05/13 15:09, Dave Scott wrote:
>>>>> On 10/05/13 14:46, George Dunlap wrote:
>>>>>> On Fri, Apr 26, 2013 at 3:29 PM, Ian Campbell <Ian.Campbell@citrix.com> wrote:
>>>>>>> On Fri, 2013-04-26 at 15:17 +0100, Sylvain Munaut wrote:
>>>>>>>>> Since the intention of that commit was to allow for qdisk backends with no
>>>>>>>>> explicit file in dom0 (i.e. network remote backend such as ceph) the lowest
>>>>>>>>> impact fix appears to be to make that explicit. This should probably be
>>>>>>>>> revisited to rationalize the probing.
>>>>>>>> What about the remote disk case of blktap ?  blktap2.5 supports NBD
>>>>>>>> already AFAIK
>>>>>>>> and I'm pretty sure I'll hit that same stat issue soon for another
>>>>>>>> remote blktap case.
>>>>>>> Right, sounds like I should have gone with my first instinct which was:
>>>>>>>
>>>>>>> 8<------------------------------------------------------------
>>>>>>>
>>>>>>>       From 884beff4a897d785f61705dcfa2f048536982d7c Mon Sep 17 00:00:00 2001
>>>>>>> From: Ian Campbell <ian.campbell@citrix.com>
>>>>>>> Date: Fri, 26 Apr 2013 12:41:43 +0100
>>>>>>> Subject: [PATCH] libxl: stat the path for all non-qdisk backends (including unknown)
>>>>>>>
>>>>>>> The commit a8a1f236a296 "libxl: Only call stat() when adding a disk if we
>>>>>>> expect a device to exist." changed things to only stat the file when the phy
>>>>>>> backend was explicitly requested. This broke the case where we are probing and
>>>>>>> would normally be able to decide on the phy option.
>>>>>> So at the moment qdisk backends aren't checked at all with this --
>>>>>> which means that if you give a path to a file that doesn't exist via,
>>>>>> for example, xl cd-insert, things fail in weird ways:
>>>>>>
>>>>>> 1. In qemu-traditional, the command silently completes; the effect is
>>>>>> that the disk currently in the drive is ejected
>>>>>>
>>>>>> 2. in qemu-upstream, qmp returns an error which is reported.  The disk
>>>>>> is ejected from the guest, but the xenstore entries are not updated
>>>>>> (still contain the old values)
>>>>>>
>>>>>> It seems like we should probably also at least check if disk_format is RAW.
>>>>>>
>>>>>> OTOH, I don't seen an option for disk_format to be ceph; is it just
>>>>>> listed as "raw" as well?  That doesn't seem right...
>>>>> AFAICT a ceph disk is in the "raw" format but it uses a custom network
>>>>> protocol to actually read and write the blocks. I imagine on the ceph
>>>>> servers the disk is stored in a cleverer format, but all qemu/tapdisk
>>>>> see are plain blocks with no fancy encoding, no .vhd or .qcow.
>>>> Oh, hang on -- does qdisk / tapdisk then just open a plain file? Then it
>>>> *is* just a raw file for these purposes; and we should still be able to
>>>> call stat() on it.
>>> There is no file :) All we have is a "URL"-like thing which looks like:
>>>
>>>       rbd:rbd/foo.img
>>>
>>> which qemu/tapdisk accesses via a C library called librados. The C
>>> library takes the "URL" and makes a network connection to a storage
>>> server and reads and writes blocks over TCP/IP. No filesystems are
>>> involved, "rbd:rbd/foo.img" doesn't exist in the filesystem.
>> Right -- so I think this should be a separate type to RAW.  I agree that
>> making a new format for each new protocol isn't very scalable -- but
>> there should be a "URL" or "VIRTUAL" format to indicate that there is no
>> real file.
> Something like that would be fine. I agree "format=raw" is a bit odd,
> when there's no real "format" involved, only a network protocol.
>
>> Right now RAW doesn't seem to be talking about the format, but about the
>> format string; i.e., "I pass this straight into qemu and it figures out
>> what it means."  That's not right.
>>
>> For 4.3 -- I guess the simplest thing to fix the actual bug (xl
>> cd-insert giving weird results) is to put a stat in xl_cmdimpl.c before
>> calling libxl_cdrom_insert().
> That works for me. Do non-CDROM cases fail in a more reasonable way?

Well presumably at some point qemu will return an error if the spec 
(either a plain file or a network thing) doesn't work properly. Whether 
that will cause a proper failure for "xl block-attach" or 
libxl_device_disk_attach() I'm not sure.

  -George