[Rd] [New Patch] Fix disk corruption when writing

[Duncan Murdoch]

On 06/07/2017 5:21 AM, Serguei Sokol wrote:
> I propose the following patch against the current R-devel/src/main/connection.c (cf. attached file).
> It gives (on my linux box):
>  > fc=file("/dev/full", "w")
>  > write.csv("a", file=fc)
> Error in writeLines(paste(col.names, collapse = sep), file, sep = eol) :
>    system call failure on writing
>  > close(fc)
>
> Serguei.

I suspect that flushing on every write will slow things down too much.

I think a better approach is to catch the error in the Rconn_printf 
calls (as R-devel currently does), and also catch errors on 
con->close(con).  This one requires more changes to the source, so it 
may be a day or two before I commit.

One thing I have to look into:  is anyone making use of the fact that 
the R-level close.connection() function can return -1 to signal an 
error?  Base R ignores that, which is why we need to do something, but 
if packages are using it, things need to be changed carefully.  I can't 
just change it to raise an error instead.

Duncan Murdoch


>
> Le 05/07/2017 à 15:33, Serguei Sokol a écrit :
>> Le 05/07/2017 à 14:46, Serguei Sokol a écrit :
>>> Le 05/07/2017 à 13:09, Duncan Murdoch a écrit :
>>>> On 05/07/2017 5:26 AM, January W. wrote:
>>>>> I tried the newest patch, but it does not seem to work for me (on
>>>>> Linux). Despite the check in Rconn_printf, the write.csv happily writes
>>>>> to /dev/full and does not report an error. When I added a printf("%d\n",
>>>>> res); to the Rconn_printf() definition, I see only positive values
>>>>> returned by the vfprintf call.
>>>>>
>>>>
>>>> That's likely because you aren't writing enough to actually trigger a write to disk during the write.  Writes are buffered, and the error doesn't happen
>>>> until the buffer is written.
>>> I can confirm this behavior with fvprintf(). Small and medium sized writings
>>> on /dev/full don't trigger error and 1MB does.
>>>
>>> But if fprintf() is used, it returns a negative value from the very first byte written.
>> I correct myself. In my test, fprintf() returned -1 for another reason (connection was already closed
>> at this moment)
>> However, if fvprintf(...) is followed by res=fflush(con) then res is -1
>> if we try to write on /dev/full. May be we have to use this to trigger
>> an error message in R.
>>
>> Serguei.
>>
>>>>   The regression test I put in had this problem; I'm working on MacOS and Windows, so I never got to actually try it before committing.
>>>>
>>>> Unfortunately, it doesn't look possible to catch the final flush of the buffer when the connection is closed, so small writes won't trigger any error.
>>>>
>>>> It's also possible that whatever system you're on doesn't signal an error when the write fails.
>>>>
>>>> Duncan Murdoch
>>>>
>>>>> Cheers,
>>>>>
>>>>> j.
>>>>>
>>>>>
>>>>> On 4 July 2017 at 21:37, Duncan Murdoch <murdoch.duncan at gmail.com
>>>>> <mailto:murdoch.duncan at gmail.com>> wrote:
>>>>>
>>>>>     On 04/07/2017 11:50 AM, Jean-Sébastien Bevilacqua wrote:
>>>>>
>>>>>         Hello,
>>>>>         You can find here a patch to fix disk corruption.
>>>>>         When your disk is full, the write function exit without error
>>>>>         but the file
>>>>>         is truncated.
>>>>>
>>>>> https://bugs.r-project.org/bugzilla/show_bug.cgi?id=17243
>>>>> <https://bugs.r-project.org/bugzilla/show_bug.cgi?id=17243>
>>>>>
>>>>>
>>>>>     Thanks.  I didn't see that when it came through (or did and forgot).
>>>>>     I'll probably move the error check to a lower level (in the
>>>>>     Rconn_printf function), if tests show that works.
>>>>>
>>>>>     Duncan Murdoch
>>>>>
>>>>>
>>>>>     ______________________________________________
>>>>>     R-devel at r-project.org <mailto:R-devel at r-project.org> mailing list
>>>>>     https://stat.ethz.ch/mailman/listinfo/r-devel
>>>>>     <https://stat.ethz.ch/mailman/listinfo/r-devel>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> -------- January Weiner --------------------------------------
>>>>
>>>> ______________________________________________
>>>> R-devel at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>>
>>>
>>
>