[PATCH v4 05/15] mm: introduce execmem_alloc() and execmem_free()

[Masami Hiramatsu (Google)]

On Sat, 20 Apr 2024 07:22:50 +0300
Mike Rapoport <rppt at kernel.org> wrote:

> On Fri, Apr 19, 2024 at 02:42:16PM -0700, Song Liu wrote:
> > On Fri, Apr 19, 2024 at 1:00 PM Mike Rapoport <rppt at kernel.org> wrote:
> > >
> > > On Fri, Apr 19, 2024 at 10:32:39AM -0700, Song Liu wrote:
> > > > On Fri, Apr 19, 2024 at 10:03 AM Mike Rapoport <rppt at kernel.org> wrote:
> > > > [...]
> > > > > > >
> > > > > > > [1] https://lore.kernel.org/all/20240411160526.2093408-1-rppt@kernel.org
> > > > > >
> > > > > > For the ROX to work, we need different users (module text, kprobe, etc.) to have
> > > > > > the same execmem_range. From [1]:
> > > > > >
> > > > > > static void *execmem_cache_alloc(struct execmem_range *range, size_t size)
> > > > > > {
> > > > > > ...
> > > > > >        p = __execmem_cache_alloc(size);
> > > > > >        if (p)
> > > > > >                return p;
> > > > > >       err = execmem_cache_populate(range, size);
> > > > > > ...
> > > > > > }
> > > > > >
> > > > > > We are calling __execmem_cache_alloc() without range. For this to work,
> > > > > > we can only call execmem_cache_alloc() with one execmem_range.
> > > > >
> > > > > Actually, on x86 this will "just work" because everything shares the same
> > > > > address space :)
> > > > >
> > > > > The 2M pages in the cache will be in the modules space, so
> > > > > __execmem_cache_alloc() will always return memory from that address space.
> > > > >
> > > > > For other architectures this indeed needs to be fixed with passing the
> > > > > range to __execmem_cache_alloc() and limiting search in the cache for that
> > > > > range.
> > > >
> > > > I think we at least need the "map to" concept (initially proposed by Thomas)
> > > > to get this work. For example, EXECMEM_BPF and EXECMEM_KPROBE
> > > > maps to EXECMEM_MODULE_TEXT, so that all these actually share
> > > > the same range.
> > >
> > > Why?
> > 
> > IIUC, we need to update __execmem_cache_alloc() to take a range pointer as
> > input. module text will use "range" for EXECMEM_MODULE_TEXT, while kprobe
> > will use "range" for EXECMEM_KPROBE. Without "map to" concept or sharing
> > the "range" object, we will have to compare different range parameters to check
> > we can share cached pages between module text and kprobe, which is not
> > efficient. Did I miss something?

Song, thanks for trying to eplain. I think I need to explain why I used
module_alloc() originally.

This depends on how kprobe features are implemented on the architecture, and
how much features are supported on kprobes.

Because kprobe jump optimization and kprobe jump-back optimization need to
use a jump instruction to jump into the trampoline and jump back from the
trampoline directly, if the architecuture jmp instruction supports +-2GB range
like x86, it needs to allocate the trampoline buffer inside such address space.
This requirement is similar to the modules (because module function needs to
call other functions in the kernel etc.), at least kprobes on x86 used
module_alloc().

However, if an architecture only supports breakpoint/trap based kprobe,
it does not need to consider whether the execmem is allocated.

> 
> We can always share large ROX pages as long as they are within the correct
> address space. The permissions for them are ROX and the alignment
> differences are due to KASAN and this is handled during allocation of the
> large page to refill the cache. __execmem_cache_alloc() only needs to limit
> the search for the address space of the range.

So I don't think EXECMEM_KPROBE always same as EXECMEM_MODULE_TEXT, it
should be configured for each arch. Especially, if it is only used for
searching parameter, it looks OK to me.

Thank you,

> 
> And regardless, they way we deal with sharing of the cache can be sorted
> out later.
> 
> > Thanks,
> > Song
> 
> -- 
> Sincerely yours,
> Mike.
> 


-- 
Masami Hiramatsu (Google) <mhiramat at kernel.org>