[PATCH v2 1/3] module: Rename module_alloc() to text_alloc() and move to kernel proper
ardb at kernel.org
Tue Jul 14 08:19:24 EDT 2020
On Tue, 14 Jul 2020 at 14:31, Peter Zijlstra <peterz at infradead.org> wrote:
> On Tue, Jul 14, 2020 at 11:28:27AM +0100, Will Deacon wrote:
> > As Ard says, module_alloc() _is_ special, in the sense that the virtual
> > memory it allocates wants to be close to the kernel text, whereas the
> > concept of allocating executable memory is broader and doesn't have these
> > restrictions. So, while I'm in favour of having a text_alloc() interface
> > that can be used by callers which only require an executable mapping, I'd
> > much prefer for the module_alloc() code to remain for, err, modules.
> So on x86 all those things (kprobes, bpf, ftrace) require that same
> An interface like the late vmalloc_exec() will simply not work for us.
Fair enough. So for x86, implementing text_alloc() as an alias of
module_alloc() makes sense. But that is not the case in general.
> We recently talked about arm64-kprobes and how you're not doing any of
> the optimizations and fully rely on the exception return. And I see
> you're one of the few archs that has bpf_jit_alloc_exec() (also,
> shouldn't you be using PAGE_KERNEL_EXEC there?). But the BPF core seems
> to use module_alloc() as a default means of allocating text.
Indeed. Which means it uses up module space which may be scarce,
especially on 32-bit ARM, and gets backed by kasan shadow pages, which
only makes sense for modules (if CONFIG_KASAN=y)
> So what should this look like? Have a text_alloc() with an argument that
> indicates where? But then I suppose we also need a means to manage PLT
> entries. Otherwise I don't exactly see how you're going to call BPF
> code, or how that BPF stuff is going to call back into its helpers.
If x86 chooses to back its implementation of text_alloc() by
module_alloc(), that is absolutely fine. But arm64 has no use for
text_alloc() at all today (bpf and kprobes don't use module_alloc(),
and ftrace does not implement dynamic trampoline allocation), and in
the general case, bpf, kprobes, ftrace and the module loader all have
different requirements that deviate subtly between architectures.
So perhaps the answer is to have text_alloc() not with a 'where'
argument but with a 'why' argument. Or more simply, just have separate
alloc/free APIs for each case, with generic versions that can be
overridden by the architecture.
More information about the linux-riscv