[CRIU] Re: [PATCH] reg-files: Serialize linking of ghost files

[Pavel Emelyanov]

On 09/17/2012 06:28 PM, Cyrill Gorcunov wrote:
> On Mon, Sep 17, 2012 at 06:22:51PM +0400, Pavel Emelyanov wrote:
>> On 09/17/2012 02:47 PM, Cyrill Gorcunov wrote:
>>> On Mon, Sep 17, 2012 at 02:31:18PM +0400, Cyrill Gorcunov wrote:
>>>> On Mon, Sep 17, 2012 at 02:26:47PM +0400, Pavel Emelyanov wrote:
>>>>> On 09/17/2012 12:50 PM, Cyrill Gorcunov wrote:
>>>>>> In case if there at least two tasks with same ghost
>>>>>> file we've a race in open_path -- both task may enter
>>>>>> the open_path at same time and try to link ghost file
>>>>>> to same name leading to
>>>>>>
>>>>>>  | 26445: Error (files-reg.c:363): Can't link
>>>>>>  | /home/crtools/test/zdtm/live/static/write_read10.test (deleted).cr.1.ghost ->
>>>>>>  | /home/crtools/test/zdtm/live/static/write_read10.test (deleted)
>>>>>>
>>>>>> The both tasks have ghost files as
>>>>>>
>>>>>>  | id:  0x3 flags: 0x8002 pos: 0000000000000000 ... name: "/home/crtools/test/zdtm/live/static/write_read10.test (deleted)"
>>>>>>
>>>>>> and
>>>>>>
>>>>>>  | id: 0x10 flags: 0x8002 pos: 0000000000000000 ... name: "/home/crtools/test/zdtm/live/static/write_read10.test (deleted)"
>>>>>>
>>>>>> Thus, to serialize link/unlink procedure we use per-ghost-file mutex.
>>>>>
>>>>> In that case users can be just unsigned int
>>>>
>>>> Yes, but I didn't want to squash everything in one patch. I'll do that on top.
>>>
>>> Here is one on top.
>>
>> There's a race with two ghosts of the same name.
> 
> Pavel, would it be an error if we just generate names for unlinked files?
> I mean instead of that "  (deleted)" name created by dcache for us we could
> simply generate unique names, no?

No. Use big-lock for this. There won't be 10000000-s of ghosts.

> 	Cyrill
> .
>