Re: [PATCH v2 02/17] mm: introduce xvmalloc() et al and use for grant table allocations

[Julien Grall <julien@xxxxxxx>]

Hi Jan,

On 26/11/2020 15:18, Jan Beulich wrote:
> On 26.11.2020 14:22, Julien Grall wrote:
> > On 26/11/2020 11:34, Jan Beulich wrote:
> > > On 25.11.2020 20:48, Stefano Stabellini wrote:
> > > > On Wed, 25 Nov 2020, Jan Beulich wrote:
> > > > > On 25.11.2020 13:15, Julien Grall wrote:
> > > > > > On 23/11/2020 14:23, Jan Beulich wrote:
> > > > > > > I'm unconvinced of the mentioning of "physically contiguous" in the
> > > > > > > comment at the top of the new header: I don't think xmalloc() provides
> > > > > > > such a guarantee. Any use assuming so would look (latently) broken to
> > > > > > > me.
> > > > > > I haven't had the chance to reply to the first version about this. So I
> > > > > > will reply here to avoid confusion.
> > > > > >
> > > > > > I can at least spot one user in Arm that would use xmalloc() that way
> > > > > > (see the allocation of itt_addr in arch/arm/gic-v3-its.c).
> > > > > And I surely wouldn't have spotted this, even if I had tried
> > > > > to find "offenders", i.e. as said before not wanting to alter
> > > > > the behavior of existing code (beyond the explicit changes
> > > > > done here) was ...
> > > > >
> > > > > > AFAIK, the memory is for the sole purpose of the ITS and should not be
> > > > > > accessed by Xen. So I think we can replace by a new version of
> > > > > > alloc_domheap_pages().
> > > > > >
> > > > > > However, I still question the usefulness of introducing yet another way
> > > > > > to allocate memory (we already have alloc_xenheap_pages(), xmalloc(),
> > > > > > alloc_domheap_pages(), vmap()) if you think users cannot rely on
> > > > > > xmalloc() to allocate memory physically contiguous.
> > > > > ... the reason to introduce a separate new interface. Plus of
> > > > > course this parallels what Linux has.
> > > > >
> > > > > > It definitely makes more difficult to figure out when to use xmalloc()
> > > > > > vs xvalloc().
> > > > > I don't see the difficulty:
> > > > > - if you need physically contiguous memory, use alloc_xen*_pages(),
> > > > > - if you know the allocation size is always no more than a page,
> > > > >     use xmalloc(),
> > If that's then intention, then may I ask why xmalloc() is able to
> > support multiple pages allocation?
> Because support for this pre-dates even the introduction of vmalloc()?
Right, so the code should disappear if we want people to avoid making 
that assumption with xmalloc().
> > Your assumption is Xen will always be built with the same page size
> > across all the architecture. While Xen only works with 4KB pages today,
> > Arm can support 16KB and 64KB. I have long term plan to add support for it.
> >
> > So I don't think you can use the page size as a way to distinguish which
> > one to use.
> The let's abstract this one level further
>
> - if you know the allocation size is always no more than the smallest
>    possible page size, use xmalloc()
So basically, xmalloc() is becoming pointless when xvmalloc() can do the 
same for you (as it would call xmalloc()).
> > > > What if you need memory physically contiguous but not necessarily an
> > > > order of pages, such as for instance 5200 bytes?
> > > This case is, I think, rare enough (in particular in Xen) that the
> > > waste of space can be tolerated imo.
> > This is quite departure from:
> >
> > commit b829a0ff5794ee5b0f96a0c872f6a4ed7b1007c7
> > Author: Jan Beulich <jbeulich@xxxxxxxx>
> > Date:   Thu Oct 13 10:03:43 2011 +0200
> >
> >       xmalloc: return unused full pages on multi-page allocations
> >
> >       Certain (boot time) allocations are relatively large (particularly when
> >       building with high NR_CPUS), but can also happen to be pretty far away
> >       from a power-of-two size. Utilize the page allocator's (other than
> >       Linux'es) capability of allowing to return space from higher-order
> >       allocations in smaller pieces to return the unused parts immediately.
> >
> >       Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> >       Acked-by: Keir Fraser <keir@xxxxxxx>
> >
> > I am curious to know what changed...
> Nothing. But even if something had, citing a 9 year old commit is
> not likely to point out any actual contradiction.
You can't say it is tolerable but in the past suggested that it was not 
(otherwise why would you hand back memory?).
Therefore I would like to understand why in the past this was not 
tolerable but now it is... What changed?
> > Anyway, what you wrote is very server focused. On Arm, we have plan to
> > run Xen on smaller hardware where wasting memory mean less usable RAM
> > for guests.
> >
> > The problem with using an order is the bigger the order is the more
> > change you will waste space...
> >
> > Allocating more than a page is fairly common on Arm, so we really want
> > to reduce the amount of memory wasted.
> The amount of space wasted is the same - the tail of the trailing
> page. I'm afraid I don't see what your point is.
There would obviously be no difference if one wants to allocate more 
than one page but less than 2 pages....
But that was not my point. My point is when you allocate with an order 
greater or equal to 2, then you will start wasting memory when not using 
xmalloc().
For instance, if you want to allocate 20kB then you will need to 
allocate 32kB and lose 12kB. To make it sound bigger, you could replace 
kB by mB.
> > > > If xmalloc can't do physically contiguous allocations, we need something
> > > > else that does physically contiguous allocations not only at page
> > > > granularity, right?
> > > Well, we first need to settle on what guarantees xmalloc() is meant
> > > to provide. It may be just me assuming it doesn't provide the same
> > > ones which Linux'es kmalloc() makes. I'm first and foremost
> > > judging by the comment near the top of xmalloc.h, which compares
> > > with malloc() / free(), not kmalloc() / kfree().
> > >
> > > > The other issue is semantics. If xmalloc is unable to allocate more than
> > > > a page of contiguous memory, then it is identical to vmalloc from the
> > > > caller's point of view: both xmalloc and vmalloc return a virtual
> > > > address for an allocation that might not be physically contiguous.
> > > Almost. vmalloc() puts guard pages around the allocation and
> > > guarantees page alignment.
> > >
> > > > Maybe we should get rid of xmalloc entirely and improve the
> > > > implementation of vmalloc so that it falls back to xmalloc for
> > > > sub-page allocations. Which in fact is almost the same thing that you
> > > > did.
> > > This would break callers assuming page alignment (and - shouldn't
> > > be an issue in practice - granularity). If anything, as Julien
> > > did suggest, we could modify xmalloc() accordingly, but then of
> > > course making sure we also honor alignment requests beyond page
> > > size.
> > >
> > > Neither of these is the goal here, hence this "intermediate"
> > > implementation, which is only almost "redundant".
> > >
> > > > > - if you know the allocation size is always more than a page, use
> > > > >     vmalloc(),
> > > > > - otherwise use xvmalloc(). Exceptions may of course apply, i.e.
> > > > > this is just a rule of thumb.
> > > > >
> > > > > > I would like to hear an opinion from the other maintainers.
> > > > > Let's hope at least one will voice theirs.
> > > > If we take a step back, I think we only really need two memory
> > > > allocators:
> > > >
> > > > 1) one that allocates physically contiguous memory
> > > > 2) one that allocates non-physically contiguous memory
> > > >
> > > > That's it, right?
> > > >
> > > > In addition to that, I understand it could be convient to have a little
> > > > wrapper that automatically chooses between 1) and 2) depending on
> > > > circumstances.
> > > >
> > > > But if the circumstance is just size < PAGE_SIZE then I don't think we
> > > > need any convenience wrappers: we should just be able to call 2), which
> > > > is vmalloc, once we improve the vmalloc implementation.
> > > >
> > > > Or do you see any reasons to keep the current vmalloc implementation as
> > > > is for sub-page allocations?
> > > See my "Almost. ..." above.
> > >
> > > As an aside, I also find it quite puzzling that in one of the rare
> > > cases where I propose to clone an interface from Linux without much
> > > deviation from their model, I get objections. It typically was the
> > > other way around in the past ...
> > If we were really following Linux, then we would have two interfaces:
> >      - xmalloc() which is the same as kmalloc()
> >      - xvalloc() which is the same a kvalloc()
> (correction: xvmalloc() and kvmalloc())
>
> - vmalloc() (named identically in Linux and Xen)
>
> IOW the same set of _three_ interface groups.
>
> > However, you seem to be the one objecting on the behavior of xmalloc().
> I don't think I'm objecting to any existing behavior. What I did
> is state my view on (non-)guarantees by xmalloc(). And I've
> already said - maybe I'm wrong and, like Linux'es kmalloc(),
> there is a guarantee of it producing contiguous memory, and I
> merely didn't find where that's said.
I can find quite a few places in Linux that use kmalloc() with size that 
are bigger than a page size.
That's enough for me to think that while this may not have been the 
original intention, people are using it like that (same in Xen). So we 
can't dismiss it.
> > I can't speak for Stefano, but I don't object on following Linux.
> > Instead I am objecting on the growing number of way to allocate memory
> > in Xen and that differ depending on the system_state.
> But as per above the addition only brings us on par with Linux.
> There, kvmalloc_node() is simply a wrapper (with different logic
> when to try what) around kmalloc_node() and __vmalloc_node(). No
> different (in the basic idea) from what I'm doing here.
There are at least two more in Xen so far:

 - alloc_domheap_pages()
 - alloc_xenheap_pages()

I still maintain that the way you suggest to use each of them is not 
clear. In particular, that xmalloc() doesn't guarantee physcally 
contiguous allocation for allocation larger than a page size.
In summary, I would be happy with the introduction of xvmalloc() as long 
as we guarantee that xmalloc() is allocating physically contiguous memory.
Users that doesn't care about this guarantee would have to be switched 
to use xvmalloc() (This doesn't need to be done here).
Cheers,

--
Julien Grall