I'd like to follow up on this a bit, because the way I found it was *weird*. MSan found an uninitialized write. Reproducing the issue through GDB showed that there's a struct mdrestore_struct type with a member of type u64 called 'leafsize' which was... half-initialized? Four bytes were uninitialized, four were initialized. Looking further, it was being set by a call to a getter function defined in ctree.h, which pulls a particular member from a superblock struct... which is of type u32. Aha! So the struct member was missized. Easy enough to fix. Except that the mdrestore_struct is initialized using calloc, so the part that wasn't getting written to was actually initialized, and the part that *was* was showing up as uninitialized. And actually contained a sane value, reliably. What was going on? Turns out that btrfs's use of zlib, which was the uninstrumented system version, caused anything allocated by uncompress() to show up as uninitialized. (Trying to use __msan_unpoison did not work, as I have an iffy install of the sanitizer in the first place (it's whatever happened to come with clang-3.3), and MSanDR (the dynamic-instrumentation tool which marks all uninstrumented libraries as good) is apparently not trivial to install. I may just build zlib.) But the point is, it would have been annoying and sad if this had been a false positive through and through, but I *did* find an actual issue here using MSan, even though MSan was pointing in the exact wrong place. Adam Buchbinder -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
