On Sun, 31 Jan 2021 at 19:55, Catalin Marinas <catalin.mari...@arm.com> wrote: > > On Fri, Jan 29, 2021 at 02:09:05PM -0800, Linus Torvalds wrote: > > On Fri, Jan 29, 2021 at 11:03 AM Catalin Marinas > > <catalin.mari...@arm.com> wrote: > > > > > > arm64 fixes: > > > > > > - Fix the virt_addr_valid() returning true for < PAGE_OFFSET addresses. > > > > That's a really odd fix. > > > > It went from an incorrect bitwise operation (masking) to an _odd_ > > bitwise operation (xor). > > > > Yes, PAGE_OFFSET has the bit pattern of all upper bits set, so "(addr > > ^ PAGE_OFFSET)" by definition reverses the upper bits - and for a > > valid case turns them to zero. > > > > But isn't the *logical* thing to do to use a subtract instead? For the > > valid cases, the two do the same thing (clear the upper bits), but > > just conceptually, isn't the operation that you actually want to do > > "(addr - PAGE_OFFSET)"? > > > > IOW, why is it using that odd xor pattern that doesn't make much > > sense? I believe it _works_, but it looks very strange to me. > > This macro used to test a single bit and it evolved into a bitmask. So, > yes, basically what we need is: > > #define __is_lm_address(addr) ((u64)(addr) >= PAGE_OFFSET && \ > (u64)(addr) < PAGE_END) > > I wasn't sure whether the code generation with two comparisons is > similar to the xor variant but the compiler should probably be smart > enough to use CMP and CCMP. In the grand scheme, it probably doesn't > even matter. > > Unless I miss something, I don't see any overflow issues even if we do > (((u64)addr - PAGE_OFFSET) < (PAGE_END - PAGE_OFFSET)). > > We can backport the fix already upstream and clean-up the code in > mainline going forward (after some sanity check on the code generation). > It would be easier to parse in the future. > > > Also, shouldn't _lm_to_phys() do the same? It does that "mask upper > > bits" too that was problematic in __is_lm_address(). Again, shouldn't > > that logically be a subtract op? > > Yes, that's similar and a subtract should do. >
The original bit test was written like that because it removes the need to reason about a potential tag in the upper bits. I tried to preserve that behavior when removing the guaranteed 1:1 split between the vmalloc and linear regions, by masking with PAGE_OFFSET and comparing with PAGE_END - PAGE_OFFSET, but unfortunately, both approaches suffer from the issue fixed by this patch, i.e., that virt_addr_valid(0x0) erroneously returns true. I think both proposed fixes are appropriate, but they both reintroduce the need to consider the tag. I don't know whether or where this could pose a problem, but it needs to be taken into account.
