Comments on: SetFileValidData Function (Windows) – Now with added FAIL

By: Sch

Sch — Tue, 28 Oct 2008 17:12:59 +0000

Have a look at the DeviceIoControl() and FSCTL_SET_ZERO_DATA.

By: MacPlusG3

MacPlusG3 — Tue, 09 Sep 2008 00:53:14 +0000

yep, XFS actually solves the problem :)

Although, in the past this led to the “null bytes in files” for apps that didn’t write their files to disk properly.

By: Brian Smith

Brian Smith — Mon, 08 Sep 2008 18:05:41 +0000

xfs has a writeback mode similar to ext3’s, but as you explained above, unwritten blocks are implicitly zero, so xfs will not have the same problem as ext3.

Some day I’d like to have a convention for all the Brian Smiths in the world (except the ones with criminal records; those guys always cause us tons of problems).

By: MacPlusG3

MacPlusG3 — Mon, 08 Sep 2008 17:47:14 +0000

They likely write each block… I really should look at recent postgresql though… another RDBMS where we can actually look at the source.

i hadn’t thought that hard about ext3 data=writeback in that way before, but yeah – not good either.

(and talk about double take.. my dad’s name is Brian Smith, and spends none of his time caring about ext3 data=writeback)

By: Brian Smith

Brian Smith — Mon, 08 Sep 2008 17:28:42 +0000

ext3 data=writeback will also expose previously-deleted data: If you append data to a file, and the system crashes after the metadata is written and before the data is written, then after the crash the end of the file will contain old deleted data.

By: LinuxJedi

LinuxJedi — Mon, 08 Sep 2008 13:37:31 +0000

And I thought I had a headache before reading that :)

I would imagine a place to start would be to look at how virtualization apps do it, as they must have a similar goal when pre-creating the virtual disks.

This has gone far beyond my Windows programming skills though :)

By: MacPlusG3

MacPlusG3 — Mon, 08 Sep 2008 13:28:39 +0000

I’m thinking about doing that, followed by the writing of zeros from start to finish… but that still leaves a point where after crash the content of the file is undefined…. Although in a RDBMS we have (or at least should) a way to roll back the not-yet-completed create of a file

By: LinuxJedi

LinuxJedi — Mon, 08 Sep 2008 13:21:51 +0000

I’ve heard you can do this with SetFilePointer() and then setting SetEndOfFile() and setting SetFilePointer() to 0 when you finish. But I have nothing to test this on right now.