Ticket #2911 (closed defect: duplicate)
bad reads on SATA disks
|Reported by:||schnittchen||Owned by:|
I had massive filesystem corruption for no apparent reason, so I investigated a bit. Apparently, reads on (guest, virtual) SATA disks can produce wrong data.
My setups were linux (Ubuntu 8.04 and 8.10) guest and host (all combinations). In order to locate the problem, I attached a vmdk disk image with prepared data so that I could tell from the read data at which blocks something had gone wrong. The image came in two variations:
- an "entire disk access" vmdk pointing to an entire mapper device
- a modified version pointing to a file, created by editing the vmdk by hand.
The problem appeared in both situations.
This is a "heisenbug", I have read many megabytes of intact data and could not reproduce the symptom until it finally hit again. It seems to appear more often when a large number of blocks is read (I did the read using dd inside the guest, so I played with the block size parameter). It seems to be very well reproducable within 150 MBytes (my prepared disk was 25MB).
There is no such problem with IDE. After detaching the vmdk and reattaching on an IDE slot it disappeared, and reappeared when on SATA again.
I captured all badly read blocks in order to get a feeling for it. I looked at the first few kilobytes and found not a single sector from the prepared disk image! The data seems to come somewhere out of RAM, or from the other attached (IDE) disk of the guest. I have seen symbols, format strings and even directory nodes.
uname of one of the host systems: Linux myhost 2.6.24-23-generic #1 SMP Thu Nov 27 18:44:42 UTC 2008 i686 GNU/Linux
The guest is almost the same. vbox version is 2.1.0
I hope this is enough information to reproduce and locate the problem.