On 01/05/2016 02:13 PM, Mike Snitzer wrote: > On Tue, Jan 05 2016 at 3:16pm -0500, > Mike Snitzer wrote: > >> On Tue, Dec 08 2015 at 1:26pm -0500, >> Scotty Bauer wrote: >> >>> Friendly ping, is anyone interested in this? >> >> The passed @user argument is flagged via __user so it can be >> deferenced directly. It does look like directly deferencing >> user->version is wrong. >> >> But even if such indirect access is needed (because __user flag is only >> applicable to @user arg, not the contained version member) we could more >> easily just do something like this no?: >> >> uint32_t __user *versionp = (uint32_t __user *)user->version; >> ... >> if (copy_from_user(version, versionp, sizeof(version))) >> return -EFAULT; >> >> I've staged the following, thanks: >> https://git.kernel.org/cgit/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-4.5&id=bffc9e237a0c3176712bcd93fc6a184a61e0df26 > > Alasdair helped me understand that we do need your original fix. > I've staged it for 4.5 (and stable@) here: > https://git.kernel.org/cgit/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-4.5&id=ead3db62bf10fe143bec99e7b7ff370d7a6d23ef > > Thanks again, > Mike > -- This broke linux-next because I'm dumb and didn't test it. I thought it was a trivial enough of a patch that I wouldn't screw it up, but I did. I incorrectly assumed that user->version was essentially a pointer in userland, not a flat chunk of memory. Ie it was a pointer to some malloc'd region, not an inlined version[3]. I thought it was this: struct dm_ioctl { uint32_t *version; ... } It is really this: struct dm_ioctl { uint32_t version[3]; } I was trying to get the values out of *version, which would have been a pointer, but instead what the code ended up doing was actually getting 8 bytes of the version (think 4,3,1) out and trying to access that version as a memory address, oops. It turns out that the original code is correct and doesn't actually touch user memory without a copy_from_user(). Gcc is smart enough to see that version[3] is inlined, and it can emit code which simply takes the userland pointer (struct dm_ioctl __user user), and calculates on offset based on the pointer, thus no actual user dereference occurs. Had the struct looked like the first example I believe the patch would work. I'm wondering now if we should switch the code a bit to make it less ambiguous, so someone like me doesn't come along again thinking the code dereferences userland memory and waste everyones time. I've attached a patch based off linux-next-20150616 which reverts my broken code but adds an & to the front of user->version so it looks like the code is doing the right thing. If I should be basing my patch off something other than linux-next let me know and I'll rewrite it, or we can just revert the old patch and ignore this one. Thanks and very sorry for the confusion and breakage.