[PATCH 00/13] mm: jit/text allocator

[Puranjay Mohan]

On Fri, Jun 2, 2023 at 8:21 PM Song Liu <song at kernel.org> wrote:
>
> On Fri, Jun 2, 2023 at 2:35 AM Mark Rutland <mark.rutland at arm.com> wrote:
> >
> > On Thu, Jun 01, 2023 at 02:14:56PM -0400, Kent Overstreet wrote:
> > > On Thu, Jun 01, 2023 at 05:12:03PM +0100, Mark Rutland wrote:
> > > > For a while I have wanted to give kprobes its own allocator so that it can work
> > > > even with CONFIG_MODULES=n, and so that it doesn't have to waste VA space in
> > > > the modules area.
> > > >
> > > > Given that, I think these should have their own allocator functions that can be
> > > > provided independently, even if those happen to use common infrastructure.
> > >
> > > How much memory can kprobes conceivably use? I think we also want to try
> > > to push back on combinatorial new allocators, if we can.
> >
> > That depends on who's using it, and how (e.g. via BPF).
> >
> > To be clear, I'm not necessarily asking for entirely different allocators, but
> > I do thinkg that we want wrappers that can at least pass distinct start+end
> > parameters to a common allocator, and for arm64's modules code I'd expect that
> > we'd keep the range falblack logic out of the common allcoator, and just call
> > it twice.
> >
> > > > > Several architectures override module_alloc() because of various
> > > > > constraints where the executable memory can be located and this causes
> > > > > additional obstacles for improvements of code allocation.
> > > > >
> > > > > This set splits code allocation from modules by introducing
> > > > > jit_text_alloc(), jit_data_alloc() and jit_free() APIs, replaces call
> > > > > sites of module_alloc() and module_memfree() with the new APIs and
> > > > > implements core text and related allocation in a central place.
> > > > >
> > > > > Instead of architecture specific overrides for module_alloc(), the
> > > > > architectures that require non-default behaviour for text allocation must
> > > > > fill jit_alloc_params structure and implement jit_alloc_arch_params() that
> > > > > returns a pointer to that structure. If an architecture does not implement
> > > > > jit_alloc_arch_params(), the defaults compatible with the current
> > > > > modules::module_alloc() are used.
> > > >
> > > > As above, I suspect that each of the callsites should probably be using common
> > > > infrastructure, but I don't think that a single jit_alloc_arch_params() makes
> > > > sense, since the parameters for each case may need to be distinct.
> > >
> > > I don't see how that follows. The whole point of function parameters is
> > > that they may be different :)
> >
> > What I mean is that jit_alloc_arch_params() tries to aggregate common
> > parameters, but they aren't actually common (e.g. the actual start+end range
> > for allocation).
> >
> > > Can you give more detail on what parameters you need? If the only extra
> > > parameter is just "does this allocation need to live close to kernel
> > > text", that's not that big of a deal.
> >
> > My thinking was that we at least need the start + end for each caller. That
> > might be it, tbh.
>
> IIUC, arm64 uses VMALLOC address space for BPF programs. The reason
> is each BPF program uses at least 64kB (one page) out of the 128MB
> address space. Puranjay Mohan (CC'ed) is working on enabling
> bpf_prog_pack for arm64. Once this work is done, multiple BPF programs
> will be able to share a page. Will this improvement remove the need to
> specify a different address range for BPF programs?

Hi,
Thanks for adding me to the conversation.

The ARM64 BPF JIT used to allocate the memory using module_alloc but it
was not optimal because BPF programs and modules were sharing the 128 MB
module region. This was fixed by
91fc957c9b1d ("arm64/bpf: don't allocate BPF JIT programs in module memory")
It created a dedicated 128 MB region set aside for BPF programs.

But 128MB could get exhausted especially where PAGE_SIZE is 64KB - one
page is needed per program. This restriction was removed by
b89ddf4cca43 ("arm64/bpf: Remove 128MB limit for BPF JIT programs")

So, currently BPF programs are using a full page from vmalloc (4 KB,
16 KB, or 64 KB).
This wastes memory and also causes iTLB pressure. Enabling bpf_prog_pack
for ARM64 would fix it. I am doing some final tests and will send the patches in
1-2 days.

Thanks,
Puranjay