On Tue, 10 Feb 2009, Shawn O. Pearce wrote: > Daniel Barkalow <barkalow@xxxxxxxxxxxx> wrote: > > > > I think maybe there's aliasing in the delta base cache? If it recycled a > > struct packed_git, the cache would come up with a cached tree at offset 12 > > of the packed_git at that address, but the pack used by that struct has > > changed. > > Yup, that must be it. > > > I don't see any reason that the situation couldn't arise where you start a > > pack, look up an object in it while it's still open but the object isn't > > in the window, cache the delta base, end that packfile, eventually start a > > packfile that gets allocated in the space that was freed, produce a new > > delta against the object at exactly the same offset of the new pack (with > > the same address as the old pack), and go on happily until you try looking > > up this last delta and pull the wrong base out of the cache. > > > > I don't see any code to flush the delta cache ever, but it's hard to get a > > new packed_git allocated at the address of a freed one, except by doing a > > lot of checkpoints in fast-import... > > *ouch*. I think you found it. > > We should dump the cached_objects table in sha1_file.c during > a checkpoint in fast-import. No, that one's keyed by sha1, and doesn't get collisions; it's the delta_base_cache that's the issue; it's keyed by struct packed_git * and offset. -Daniel *This .sig left intentionally blank* -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html