Date:   Thu, 30 Jul 2020 10:09:02 +0900
From:   Masami Hiramatsu <mhiramat@...nel.org>
To:     Ard Biesheuvel <ardb@...nel.org>
Cc:     Mike Rapoport <rppt@...nel.org>,
        Jarkko Sakkinen <jarkko.sakkinen@...ux.intel.com>,
        Ingo Molnar <mingo@...nel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-mm@...ck.org, Andi Kleen <ak@...ux.intel.com>,
        Peter Zijlstra <peterz@...radead.org>,
        "Naveen N. Rao" <naveen.n.rao@...ux.ibm.com>,
        Anil S Keshavamurthy <anil.s.keshavamurthy@...el.com>,
        "David S. Miller" <davem@...emloft.net>,
        Jessica Yu <jeyu@...nel.org>
Subject: Re: [PATCH v5 5/6] kprobes: Use text_alloc() and text_free()

On Wed, 29 Jul 2020 09:13:21 +0300
Ard Biesheuvel <ardb@...nel.org> wrote:

> On Wed, 29 Jul 2020 at 04:51, Masami Hiramatsu <mhiramat@...nel.org> wrote:
> >
> > On Tue, 28 Jul 2020 20:51:08 +0300
> > Ard Biesheuvel <ardb@...nel.org> wrote:
> >
> > > On Tue, 28 Jul 2020 at 16:35, Masami Hiramatsu <mhiramat@...nel.org> wrote:
> > > >
> > > > On Tue, 28 Jul 2020 13:56:43 +0300
> > > > Ard Biesheuvel <ardb@...nel.org> wrote:
> > > >
> > > > > On Tue, 28 Jul 2020 at 11:17, Masami Hiramatsu <mhiramat@...nel.org> wrote:
> > > > > > > Masami or Peter should correct me if I am wrong, but it seems to me
> > > > > > > that the way kprobes uses these pages does not require them to be in
> > > > > > > relative branching range of the core kernel on any architecture, given
> > > > > > > that they are populated with individual instruction opcodes that are
> > > > > > > executed in single step mode, and relative branches are emulated (when
> > > > > > > needed)
> > > > > >
> > > > > > Actually, x86 and arm has the "relative branching range" requirements
> > > > > > for the jump optimized kprobes. For the other architectures, I think
> > > > > > we don't need it. Only executable text buffer is needed.
> > > > > >
> > > > >
> > > > > Thanks for the explanation. Today, arm64 uses the definition below.
> > > > >
> > > > > void *alloc_insn_page(void)
> > > > > {
> > > > >   return __vmalloc_node_range(PAGE_SIZE, 1, VMALLOC_START, VMALLOC_END,
> > > > >     GFP_KERNEL, PAGE_KERNEL_ROX, VM_FLUSH_RESET_PERMS,
> > > > >     NUMA_NO_NODE, __builtin_return_address(0));
> > > > > }
> > > > >
> > > > > Do you think we could use that as the generic implementation if we use
> > > > > MODULES_START/_END as the allocation window?
> > > >
> > > > Yes, but for the generic implementation, we don't need to consider the
> > > > relative branching range since we can override it for x86 and arm.
> > > > (and that will be almost same as module_alloc() default code)
> > >
> > > Indeed. So having kprobes specific macros that default to
> > > VMALLOC_START/END but can be overridden would be sufficient.
> > >
> > > > BTW, is PAGE_KERNEL_ROX flag available generically?
> > > >
> > >
> > > Turns out that it is not :-(
> >
> > Hmm, in that case, we need to use PAGE_KERNEL_EXEC.
> >
> > In the result, may it be similar to this? :)
> >
> > void * __weak module_alloc(unsigned long size)
> > {
> >         return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
> >                         GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
> >                         NUMA_NO_NODE, __builtin_return_address(0));
> > }
> >
> > The major difference between module_alloc() and kprobe's alloc_page_insn()
> > is the alloc_page_insn() makes the page ROX after allocating the pages *ONLY*
> > on x86 and arm64.
> >
> 
> Right.
> 
> > $ git grep -w alloc_insn_page -- arch
> > arch/arm64/kernel/probes/kprobes.c:void *alloc_insn_page(void)
> > arch/x86/kernel/kprobes/core.c:void *alloc_insn_page(void)
> >
> > However since the module_alloc() owns its arch-dependent implementations
> > most of major architectures, if we implement independent text_alloc_kprobe(),
> > we need to make deadcopies of module_alloc() for each architecture.
> >
> 
> No, that is what we are trying to avoid.
> 
> > $ git grep 'module_alloc(unsigned' arch/
> > arch/arm/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/arm64/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/mips/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/nds32/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/nios2/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/parisc/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/riscv/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/s390/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/sparc/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/unicore32/kernel/module.c:void *module_alloc(unsigned long size)
> > arch/x86/kernel/module.c:void *module_alloc(unsigned long size)
> >
> > It seems that some constrains for module_alloc() exists for above
> > architectures.
> >
> > Anyway, for kprobe's text_alloc() requirements are
> > - It must be executable for the arch which uses a single-step out-of-line.
> >   (and need to be registered to KASAN?)
> 
> No, kasan shadow is not needed here.

OK, I just saw the KASAN shadow was generated (kasan_module_alloc was
called) in module_alloc() on x86.
So current x86 alloc_insn_page() has it.

> > - It must be ROX if implemented (currently only for x86 and arm64)
> 
> x86 does not actually define thr macro, but the result is the same.

Yes, alloc_insn_page() for x86 is adding the RO and X flag on the page.

> > - It must be in the range of relative branching only for x86 and arm.
> >
> 
> So in summary, the generic module_alloc() above can be reused for
> kprobes on all arches except x86 and arm64, right? Then we can remove
> the call to it, and drop the modules dependency.

Yes, that's correct. If there is a generic module_alloc(), kprobes is
happy to use it :)

Thank you,

-- 
Masami Hiramatsu <mhiramat@...nel.org>

Powered by blists - more mailing lists