[poppler] UniqueFileStream and deleted files

[Thomas Freitag]

Am 25.03.2013 00:51, schrieb Albert Astals Cid:
> El Diumenge, 24 de març de 2013, a les 16:31:05, Adam Reichold va escriure:
>> Hello,
>>
>> Am 24.03.2013 06:32, schrieb Thomas Freitag:
>>> Am 23.03.2013 20:08, schrieb Albert Astals Cid:
>>>> El Dissabte, 23 de març de 2013, a les 20:01:58, Thomas Freitag va
>>>>
>>>> escriure:
>>>>> Am 23.03.2013 18:59, schrieb Ihar `Philips` Filipau:
>>>>>> On 3/23/13, Ihar `Philips` Filipau <thephilips at gmail.com> wrote:
>>>>>>>> (i tried to find a way to duplicate a FILE* but failed)
>>>>>>> How did you duplicate FILE*?
>>>>>>>
>>>>>>> How did the `fdopen( fileno(oldfile), mode )` failed?
>>>>>> Nope. This is the right way:
>>>>>>
>>>>>> int new_fd = dup( fileno(oldfile) );
>>>>>> FILE *new_file = fdopen( new_fd, mode );
>>>>> Duplicating the FILE pointer in that way is not a solution: I tried it
>>>>> that way when I began to implement thread safe feature: At least under
>>>>> Ubuntu and gcc the duplicated file pointer uses the same underlying
>>>>> buffer, and that corrupts the thread safe feature.
>>>> Exactly, it is the documented behaviour
>>>>
>>>> "They refer to the same open file description and thus share file offset"
>>>>
>>>>> That's why I used the
>>>>> fileName to create a complete new file pointer.
>>>>> Of course we can do it in that way if we detect that the file is deleted
>>>>> (and only then!), but a program which use threads to render different
>>>>> pages the same time will then render garbage :-(
>>>> No. We either write the file back to a temporary location as i
>>>> suggested in my
>>>> initial mail, or use the only fd with locking. I think i'm favoring
>>>> the second
>>>> use case since writing the file back might not even work if we run out
>>>> of free
>>>> space.
>>> I think, You're second solution is quite easy to implement (I mean
>>> really only easy to implement, it will take some time to change every
>>> code snippet and deeply test the changed code): since FileStream already
>>> use it's own buffer, there is no really need of the underlying system
>>> file buffer. So implementing something like a sharable GooFile (or
>>> extend gfile) which can be locked and remember the next read position in
>>> FileStream and then do something like gfile->read(readposition, buf,
>>> buflength) wouldn't really reduce the speed. But then I would prefer use
>>> the low level file handles instead of the file pointer. And it's even
>>> better than to write platform specific code and rely on Windows that the
>>> open file is not deletable and search for a Mac solution.
>>> Then the UniqueFileStream mechanism and the streamOwner instance
>>> variable can be removed, too.
>>> On the other hand: who deletes a just opened file and expects stability
>>> of the application which just opens the file? Is it really worth the
>>> effort?
>> Seems I am a bit to the party, but for what it's worth: I asked myself
>> the same question as Thomas. Providing a file object that takes care of
>> looking and per-thread offset seems like the best solution (ideally
>> implementing using low-level I/O), but is it really that useful to be
>> able to continue rendering unlinked files as long as Poppler fails
>> gracefully, i.e. does not crash.
> We still have the file open, it should work.
Ok, ok, I surrender. If You just post it to find a mug who will do it, 
then open a bug report and assign it to me :-) Even if I don't feel like 
to open an issue again which I just closed after working on it round 
about 9 months, it is easier to make the mistakes by my own instead of 
searching errors from others :-)
But I'm too busy in the moment with other projects to do it immediately, 
so it will take probably a while.

Cheers,
Thomas
>
> Cheers,
>    Albert
>
>> Best regards, Adam.
>>
>>> Cheers,
>>> Thomas
>>>
>>>> Cheers,
>>>>
>>>>     Albert
>>>>> Cheers,
>>>>> Thomas
>>>>>
>>>>>> It is OK to use *NIX function here - the dup() - since deleting open
>>>>>> file can happen only on the *NIX-like OS, Mac OS X included. Windows
>>>>>> doesn't allow that. Correct me if I'm wrong. Tested on Linux, HP-UX
>>>>>> and Solaris for the safety sake: the fclose() would close the dup()ed
>>>>>> file descriptor.
>>>>>>
>>>>>> Though I'm not sure how to integrate that with the rest of the portable
>>>>>> code. :)
>>>>>>
>>>>>> FYI.
>>>>>> _______________________________________________
>>>>>> poppler mailing list
>>>>>> poppler at lists.freedesktop.org
>>>>>> http://lists.freedesktop.org/mailman/listinfo/poppler
>>>>>>
>>>>>> .
>>>>> _______________________________________________
>>>>> poppler mailing list
>>>>> poppler at lists.freedesktop.org
>>>>> http://lists.freedesktop.org/mailman/listinfo/poppler
>>>> _______________________________________________
>>>> poppler mailing list
>>>> poppler at lists.freedesktop.org
>>>> http://lists.freedesktop.org/mailman/listinfo/poppler
>>>>
>>>> .
>>> _______________________________________________
>>> poppler mailing list
>>> poppler at lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/poppler
>> _______________________________________________
>> poppler mailing list
>> poppler at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/poppler
> _______________________________________________
> poppler mailing list
> poppler at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/poppler
>
> .
>