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

[Mike Rapoport]

On Mon, Jun 05, 2023 at 11:09:34AM +0100, Mark Rutland wrote:
> On Mon, Jun 05, 2023 at 12:20:40PM +0300, Mike Rapoport wrote:
> > On Fri, Jun 02, 2023 at 10:35:09AM +0100, Mark Rutland 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).
> > 
> > jit_alloc_arch_params() tries to aggregate architecture constraints and
> > requirements for allocations of executable memory and this exactly what
> > the first 6 patches of this set do.
> > 
> > A while ago Thomas suggested to use a structure that parametrizes
> > architecture constraints by the memory type used in modules [1] and Song
> > implemented the infrastructure for it and x86 part [2].
> > 
> > I liked the idea of defining parameters in a single structure, but I
> > thought that approaching the problem from the arch side rather than from
> > modules perspective will be better starting point, hence these patches.
> > 
> > I don't see a fundamental reason why a single structure cannot describe
> > what is needed for different code allocation cases, be it modules, kprobes
> > or bpf. There is of course an assumption that the core allocations will be
> > the same for all the users, and it seems to me that something like 
> > 
> > * allocate physical memory if allocator caches are empty
> > * map it in vmalloc or modules address space
> > * return memory from the allocator cache to the caller
> > 
> > will work for all usecases.
> > 
> > We might need separate caches for different cases on different
> > architectures, and a way to specify what cache should be used in the
> > allocator API, but that does not contradict a single structure for arch
> > specific parameters, but only makes it more elaborate, e.g. something like
> > 
> > enum jit_type {
> > 	JIT_MODULES_TEXT,
> > 	JIT_MODULES_DATA,
> > 	JIT_KPROBES,
> > 	JIT_FTRACE,
> > 	JIT_BPF,
> > 	JIT_TYPE_MAX,
> > };
> > 
> > struct jit_alloc_params {
> > 	struct jit_range	ranges[JIT_TYPE_MAX];
> > 	/* ... */
> > };
> > 
> > > > 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.
> > 
> > Do you mean that modules will have something like
> > 
> > 	jit_text_alloc(size, MODULES_START, MODULES_END);
> > 
> > and kprobes will have
> > 
> > 	jit_text_alloc(size, KPROBES_START, KPROBES_END);
> > ?
> 
> Yes.
> 
> > It sill can be achieved with a single jit_alloc_arch_params(), just by
> > adding enum jit_type parameter to jit_text_alloc().
> 
> That feels backwards to me; it centralizes a bunch of information about
> distinct users to be able to shove that into a static array, when the callsites
> can pass that information. 
> 
> What's *actually* common after separating out the ranges? Is it just the
> permissions?

Even if for some architecture the only common thing are the permissions,
having a definition for code allocations in a single place an improvement.
The diffstat of the patches is indeed positive (even without comments), but
having a single structure that specifies how the code should be allocated
would IMHO actually reduce the maintenance burden.

And features like caching of large pages and sub-page size allocations are
surely will be easier to opt-in this way.
 
> If we want this to be able to share allocations and so on, why can't we do this
> like a kmem_cache, and have the callsite pass a pointer to the allocator data?
> That would make it easy for callsites to share an allocator or use a distinct
> one.

I've looked into doing this like a kmem_cache with call sites passing the
allocator data, and this gets really hairy. For each user we need to pass
the arch specific parameters to that user, create a cache there and only
then the cache can be used. Since we don't have hooks to setup any of the
users in the arch code, the initialization gets more complex than shoving
everything into an array.

I think that jit_alloc(type, size) is the best way to move forward to let
different users choose their ranges and potentially caches. Differentiation
by the API name will explode even now and it'll get worse if/when new users
will show up and we can't even force users to avoid using PC-relative
addressing because, e.g. RISC-V explicitly switched their BPF JIT to use
that.
 
> Thanks,
> Mark.

-- 
Sincerely yours,
Mike.