Re: [RFC PATCH 1/1] mm/percpu.c: fix memory leakage issue when allocate a odd alignment area

From: Tejun Heo
Date: Thu Oct 13 2016 - 19:31:58 EST


On Tue, Oct 11, 2016 at 09:24:50PM +0800, zijun_hu wrote:
> From: zijun_hu <zijun_hu@xxxxxxx>
>
> the LSB of a chunk->map element is used for free/in-use flag of a area
> and the other bits for offset, the sufficient and necessary condition of
> this usage is that both size and alignment of a area must be even numbers
> however, pcpu_alloc() doesn't force its @align parameter a even number
> explicitly, so a odd @align maybe causes a series of errors, see below
> example for concrete descriptions.
>
> lets assume area [16, 36) is free but its previous one is in-use, we want
> to allocate a @size == 8 and @align == 7 area. the larger area [16, 36) is
> split to three areas [16, 21), [21, 29), [29, 36) eventually. however, due
> to the usage for a chunk->map element, the actual offset of the aim area
> [21, 29) is 21 but is recorded in relevant element as 20; moreover the
> residual tail free area [29, 36) is mistook as in-use and is lost silently
>
> as explained above, inaccurate either offset or free/in-use state of
> a area is recorded into relevant chunk->map element if request a odd
> alignment area, and so causes memory leakage issue
>
> fix it by forcing the @align of a area to allocate a even number
> as do for @size.
>
> BTW, macro ALIGN() within pcpu_fit_in_area() is replaced by roundup() too
> due to back reason. in order to align a value @v up to @a boundary, macro
> roundup(v, a) is more generic than ALIGN(x, a); the latter doesn't work
> well when @a isn't a power of 2 value. for example, roundup(10, 6) == 12
> but ALIGN(10, 6) == 10, the former result is desired obviously
>
> Signed-off-by: zijun_hu <zijun_hu@xxxxxxx>

Nacked-by: Tejun Heo <tj@xxxxxxxxxx>

This is a fix for an imaginary problem. The most we should do about
odd alignment is triggering a WARN_ON.

Thanks.

--
tejun