Re: [PATCH v2 2/2] mm, sl[aou]b: guarantee natural alignment for kmalloc(power-of-two)

[Roman Gushchin <guro@fb.com>]

On Mon, Sep 23, 2019 at 06:36:32PM +0200, Vlastimil Babka wrote:
> On 8/26/19 1:16 PM, Vlastimil Babka wrote:
> > In most configurations, kmalloc() happens to return naturally aligned (i.e.
> > aligned to the block size itself) blocks for power of two sizes. That means
> > some kmalloc() users might unknowingly rely on that alignment, until stuff
> > breaks when the kernel is built with e.g.  CONFIG_SLUB_DEBUG or CONFIG_SLOB,
> > and blocks stop being aligned. Then developers have to devise workaround such
> > as own kmem caches with specified alignment [1], which is not always practical,
> > as recently evidenced in [2].
> > 
> > The topic has been discussed at LSF/MM 2019 [3]. Adding a 'kmalloc_aligned()'
> > variant would not help with code unknowingly relying on the implicit alignment.
> > For slab implementations it would either require creating more kmalloc caches,
> > or allocate a larger size and only give back part of it. That would be
> > wasteful, especially with a generic alignment parameter (in contrast with a
> > fixed alignment to size).
> > 
> > Ideally we should provide to mm users what they need without difficult
> > workarounds or own reimplementations, so let's make the kmalloc() alignment to
> > size explicitly guaranteed for power-of-two sizes under all configurations.
> > What this means for the three available allocators?
> > 
> > * SLAB object layout happens to be mostly unchanged by the patch. The
> >   implicitly provided alignment could be compromised with CONFIG_DEBUG_SLAB due
> >   to redzoning, however SLAB disables redzoning for caches with alignment
> >   larger than unsigned long long. Practically on at least x86 this includes
> >   kmalloc caches as they use cache line alignment, which is larger than that.
> >   Still, this patch ensures alignment on all arches and cache sizes.
> > 
> > * SLUB layout is also unchanged unless redzoning is enabled through
> >   CONFIG_SLUB_DEBUG and boot parameter for the particular kmalloc cache. With
> >   this patch, explicit alignment is guaranteed with redzoning as well. This
> >   will result in more memory being wasted, but that should be acceptable in a
> >   debugging scenario.
> > 
> > * SLOB has no implicit alignment so this patch adds it explicitly for
> >   kmalloc(). The potential downside is increased fragmentation. While
> >   pathological allocation scenarios are certainly possible, in my testing,
> >   after booting a x86_64 kernel+userspace with virtme, around 16MB memory
> >   was consumed by slab pages both before and after the patch, with difference
> >   in the noise.
> > 
> > [1] https://lore.kernel.org/linux-btrfs/c3157c8e8e0e7588312b40c853f65c02fe6c957a.1566399731.git.christophe.leroy@c-s.fr/
> > [2] https://lore.kernel.org/linux-fsdevel/20190225040904.5557-1-ming.lei@redhat.com/
> > [3] https://urldefense.proofpoint.com/v2/url?u=https-3A__lwn.net_Articles_787740_&d=DwICaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=jJYgtDM7QT-W-Fz_d29HYQ&m=UUX1YoPTOOycNowHuRP2ZnqwSwZFjAFrkQFrqstidZ0&s=Kt_XTKlh2qxbC_7ME44MV3_QWFVRHlI1p2EQZYP0uqY&e= 
> > 
> > Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
> 
> So if anyone thinks this is a good idea, please express it (preferably
> in a formal way such as Acked-by), otherwise it seems the patch will be
> dropped (due to a private NACK, apparently).
> 
> Otherwise I don't think there can be an objective conclusion. On the one
> hand we avoid further problems and workarounds due to misalignment (or
> objects allocated beyond page boundary, which was only recently
> mentioned), on the other hand we potentially make future changes to
> SLAB/SLUB or hypotetical new implementation either more complicated, or
> less effective due to extra fragmentation. Different people can have
> different opinions on what's more important.
> 
> Let me however explain why I think we don't have to fear the future
> implementation complications that much. There was an argument IIRC that
> extra non-debug metadata could start to be prepended/appended to an
> object in the future (i.e. RCU freeing head?).
> 
> 1) Caches can be already created with explicit alignment, so a naive
> pre/appending implementation would already waste memory on such caches.
> 2) Even without explicit alignment, a single slab cache for 512k objects
> with few bytes added to each object would waste almost 512k as the
> objects wouldn't fit precisely in an (order-X) page. The percentage
> wasted depends on X.
> 3) Roman recently posted a patchset [1] that basically adds a cgroup
> pointer to each object. The implementation doesn't append it to objects
> naively however, but adds a separately allocated array. Alignment is
> thus unchanged.

To be fair here, we *might* want to put this pointer just after/before
the object to reduce the number of cache misses. I've put it into
a separate place mostly for simplicity reasons.

It's not an objection though, just a note.

Thanks!

> 
> [1] https://lore.kernel.org/linux-mm/20190905214553.1643060-1-guro@fb.com/
>