On Wed, Mar 23, 2022 at 11:34:59AM +0900, Masami Hiramatsu wrote:
> Add rethook for x86 implementation. Most of the code has been copied from
> kretprobes on x86.
Right; as said, I'm really unhappy with growing a carbon copy of this
stuff instead of sharing. Can we *please* keep it a single instance?
Them being basically indentical, it should be trivial to have
CONFIG_KPROBE_ON_RETHOOK (or somesuch) and just share this.
Also, what's rethook for anyway?
> diff --git a/arch/x86/kernel/kprobes/common.h b/arch/x86/kernel/kprobes/common.h
> index 7d3a2e2daf01..c993521d4933 100644
> --- a/arch/x86/kernel/kprobes/common.h
> +++ b/arch/x86/kernel/kprobes/common.h
> @@ -6,6 +6,7 @@
>
> #include <asm/asm.h>
> #include <asm/frame.h>
> +#include <asm/insn.h>
>
> #ifdef CONFIG_X86_64
>
> diff --git a/arch/x86/kernel/rethook.c b/arch/x86/kernel/rethook.c
> new file mode 100644
> index 000000000000..3e916361c33b
> --- /dev/null
> +++ b/arch/x86/kernel/rethook.c
> @@ -0,0 +1,121 @@
> +// SPDX-License-Identifier: GPL-2.0-or-later
> +/*
> + * x86 implementation of rethook. Mostly copied from arch/x86/kernel/kprobes/core.c.
> + */
> +#include <linux/bug.h>
> +#include <linux/rethook.h>
> +#include <linux/kprobes.h>
> +#include <linux/objtool.h>
> +
> +#include "kprobes/common.h"
> +
> +__visible void arch_rethook_trampoline_callback(struct pt_regs *regs);
> +
> +/*
> + * When a target function returns, this code saves registers and calls
> + * arch_rethook_trampoline_callback(), which calls the rethook handler.
> + */
> +asm(
> + ".text\n"
> + ".global arch_rethook_trampoline\n"
> + ".type arch_rethook_trampoline, @function\n"
> + "arch_rethook_trampoline:\n"
> +#ifdef CONFIG_X86_64
> + ANNOTATE_NOENDBR /* This is only jumped from ret instruction */
> + /* Push a fake return address to tell the unwinder it's a kretprobe. */
> + " pushq $arch_rethook_trampoline\n"
> + UNWIND_HINT_FUNC
" pushq $" __stringify(__KERNEL_DS) "\n" /* %ss */
/* Save the 'sp - 16', this will be fixed later. */
> + " pushq %rsp\n"
> + " pushfq\n"
> + SAVE_REGS_STRING
> + " movq %rsp, %rdi\n"
> + " call arch_rethook_trampoline_callback\n"
> + RESTORE_REGS_STRING
/* In the callback function, 'regs->flags' is copied to 'regs->ss'. */
this comment could do with a 'why' though... Because neither
this nor the one in the handler really explains why it is
important to have popf last
" addq $16, %rsp\n"
> + " popfq\n"
> +#else
same for i386:
> + /* Push a fake return address to tell the unwinder it's a kretprobe. */
> + " pushl $arch_rethook_trampoline\n"
> + UNWIND_HINT_FUNC
/* Save the 'sp - 8', this will be fixed later. */
" pushl %ss\n"
> + " pushl %esp\n"
> + " pushfl\n"
> + SAVE_REGS_STRING
> + " movl %esp, %eax\n"
> + " call arch_rethook_trampoline_callback\n"
> + RESTORE_REGS_STRING
/* In the callback function, 'regs->flags' is copied to 'regs->ss'. */
" addl $8, %esp\n"
> + " popfl\n"
> +#endif
> + ASM_RET
> + ".size arch_rethook_trampoline, .-arch_rethook_trampoline\n"
> +);
> +NOKPROBE_SYMBOL(arch_rethook_trampoline);
> +
> +/*
> + * Called from arch_rethook_trampoline
> + */
> +__used __visible void arch_rethook_trampoline_callback(struct pt_regs *regs)
> +{
> + unsigned long *frame_pointer;
> +
> + /* fixup registers */
> + regs->cs = __KERNEL_CS;
> +#ifdef CONFIG_X86_32
> + regs->gs = 0;
> +#endif
> + regs->ip = (unsigned long)&arch_rethook_trampoline;
> + regs->orig_ax = ~0UL;
regs->sp += 2*sizeof(long);
> + frame_pointer = ®s->sp + 1;
> +
> + /*
> + * The return address at 'frame_pointer' is recovered by the
> + * arch_rethook_fixup_return() which called from this
> + * rethook_trampoline_handler().
> + */
> + rethook_trampoline_handler(regs, (unsigned long)frame_pointer);
> +
> + /*
> + * Copy FLAGS to 'pt_regs::sp' so that arch_rethook_trapmoline()
> + * can do RET right after POPF.
> + */
regs->ss = regs->flags;
> +}
> +NOKPROBE_SYMBOL(arch_rethook_trampoline_callback);
On Wed, 23 Mar 2022 09:05:26 +0100
Peter Zijlstra <[email protected]> wrote:
> On Wed, Mar 23, 2022 at 11:34:59AM +0900, Masami Hiramatsu wrote:
> > Add rethook for x86 implementation. Most of the code has been copied from
> > kretprobes on x86.
>
> Right; as said, I'm really unhappy with growing a carbon copy of this
> stuff instead of sharing. Can we *please* keep it a single instance?
OK, then let me update the kprobe side too.
> Them being basically indentical, it should be trivial to have
> CONFIG_KPROBE_ON_RETHOOK (or somesuch) and just share this.
Yes, ideally it should use CONFIG_HAVE_RETHOOK since the rethook arch port
must be a copy of the kretprobe implementation. But for safety, I think
having CONFIG_KPROBE_ON_RETHOOK is a good idea until replacing all kretprobe
implementations.
>
> Also, what's rethook for anyway?
Rethook is a feature which hooks the function return. Most of the
logic came from the kretprobe. Simply to say, 'kretprobe - kprobe' is
the rethook :)
Thank you,
>
> > diff --git a/arch/x86/kernel/kprobes/common.h b/arch/x86/kernel/kprobes/common.h
> > index 7d3a2e2daf01..c993521d4933 100644
> > --- a/arch/x86/kernel/kprobes/common.h
> > +++ b/arch/x86/kernel/kprobes/common.h
> > @@ -6,6 +6,7 @@
> >
> > #include <asm/asm.h>
> > #include <asm/frame.h>
> > +#include <asm/insn.h>
> >
> > #ifdef CONFIG_X86_64
> >
> > diff --git a/arch/x86/kernel/rethook.c b/arch/x86/kernel/rethook.c
> > new file mode 100644
> > index 000000000000..3e916361c33b
> > --- /dev/null
> > +++ b/arch/x86/kernel/rethook.c
> > @@ -0,0 +1,121 @@
> > +// SPDX-License-Identifier: GPL-2.0-or-later
> > +/*
> > + * x86 implementation of rethook. Mostly copied from arch/x86/kernel/kprobes/core.c.
> > + */
> > +#include <linux/bug.h>
> > +#include <linux/rethook.h>
> > +#include <linux/kprobes.h>
> > +#include <linux/objtool.h>
> > +
> > +#include "kprobes/common.h"
> > +
> > +__visible void arch_rethook_trampoline_callback(struct pt_regs *regs);
> > +
> > +/*
> > + * When a target function returns, this code saves registers and calls
> > + * arch_rethook_trampoline_callback(), which calls the rethook handler.
> > + */
> > +asm(
> > + ".text\n"
> > + ".global arch_rethook_trampoline\n"
> > + ".type arch_rethook_trampoline, @function\n"
> > + "arch_rethook_trampoline:\n"
> > +#ifdef CONFIG_X86_64
> > + ANNOTATE_NOENDBR /* This is only jumped from ret instruction */
> > + /* Push a fake return address to tell the unwinder it's a kretprobe. */
> > + " pushq $arch_rethook_trampoline\n"
> > + UNWIND_HINT_FUNC
> " pushq $" __stringify(__KERNEL_DS) "\n" /* %ss */
> /* Save the 'sp - 16', this will be fixed later. */
> > + " pushq %rsp\n"
> > + " pushfq\n"
> > + SAVE_REGS_STRING
> > + " movq %rsp, %rdi\n"
> > + " call arch_rethook_trampoline_callback\n"
> > + RESTORE_REGS_STRING
> /* In the callback function, 'regs->flags' is copied to 'regs->ss'. */
>
> this comment could do with a 'why' though... Because neither
> this nor the one in the handler really explains why it is
> important to have popf last
>
> " addq $16, %rsp\n"
> > + " popfq\n"
> > +#else
>
> same for i386:
>
> > + /* Push a fake return address to tell the unwinder it's a kretprobe. */
> > + " pushl $arch_rethook_trampoline\n"
> > + UNWIND_HINT_FUNC
> /* Save the 'sp - 8', this will be fixed later. */
> " pushl %ss\n"
> > + " pushl %esp\n"
> > + " pushfl\n"
> > + SAVE_REGS_STRING
> > + " movl %esp, %eax\n"
> > + " call arch_rethook_trampoline_callback\n"
> > + RESTORE_REGS_STRING
> /* In the callback function, 'regs->flags' is copied to 'regs->ss'. */
> " addl $8, %esp\n"
> > + " popfl\n"
> > +#endif
> > + ASM_RET
> > + ".size arch_rethook_trampoline, .-arch_rethook_trampoline\n"
> > +);
> > +NOKPROBE_SYMBOL(arch_rethook_trampoline);
> > +
> > +/*
> > + * Called from arch_rethook_trampoline
> > + */
> > +__used __visible void arch_rethook_trampoline_callback(struct pt_regs *regs)
> > +{
> > + unsigned long *frame_pointer;
> > +
> > + /* fixup registers */
> > + regs->cs = __KERNEL_CS;
> > +#ifdef CONFIG_X86_32
> > + regs->gs = 0;
> > +#endif
> > + regs->ip = (unsigned long)&arch_rethook_trampoline;
> > + regs->orig_ax = ~0UL;
> regs->sp += 2*sizeof(long);
> > + frame_pointer = ®s->sp + 1;
> > +
> > + /*
> > + * The return address at 'frame_pointer' is recovered by the
> > + * arch_rethook_fixup_return() which called from this
> > + * rethook_trampoline_handler().
> > + */
> > + rethook_trampoline_handler(regs, (unsigned long)frame_pointer);
> > +
> > + /*
> > + * Copy FLAGS to 'pt_regs::sp' so that arch_rethook_trapmoline()
> > + * can do RET right after POPF.
> > + */
> regs->ss = regs->flags;
> > +}
> > +NOKPROBE_SYMBOL(arch_rethook_trampoline_callback);
--
Masami Hiramatsu <[email protected]>
On Wed, Mar 23, 2022 at 4:41 AM Masami Hiramatsu <[email protected]> wrote:
>
> On Wed, 23 Mar 2022 09:05:26 +0100
> Peter Zijlstra <[email protected]> wrote:
>
> > On Wed, Mar 23, 2022 at 11:34:59AM +0900, Masami Hiramatsu wrote:
> > > Add rethook for x86 implementation. Most of the code has been copied from
> > > kretprobes on x86.
> >
> > Right; as said, I'm really unhappy with growing a carbon copy of this
> > stuff instead of sharing. Can we *please* keep it a single instance?
>
> OK, then let me update the kprobe side too.
>
> > Them being basically indentical, it should be trivial to have
> > CONFIG_KPROBE_ON_RETHOOK (or somesuch) and just share this.
>
> Yes, ideally it should use CONFIG_HAVE_RETHOOK since the rethook arch port
> must be a copy of the kretprobe implementation. But for safety, I think
> having CONFIG_KPROBE_ON_RETHOOK is a good idea until replacing all kretprobe
> implementations.
Masami,
you're respinning this patch to combine
arch_rethook_trampoline and __kretprobe_trampoline
right?
On Thu, 24 Mar 2022 19:03:43 -0700
Alexei Starovoitov <[email protected]> wrote:
> On Wed, Mar 23, 2022 at 4:41 AM Masami Hiramatsu <[email protected]> wrote:
> >
> > On Wed, 23 Mar 2022 09:05:26 +0100
> > Peter Zijlstra <[email protected]> wrote:
> >
> > > On Wed, Mar 23, 2022 at 11:34:59AM +0900, Masami Hiramatsu wrote:
> > > > Add rethook for x86 implementation. Most of the code has been copied from
> > > > kretprobes on x86.
> > >
> > > Right; as said, I'm really unhappy with growing a carbon copy of this
> > > stuff instead of sharing. Can we *please* keep it a single instance?
> >
> > OK, then let me update the kprobe side too.
> >
> > > Them being basically indentical, it should be trivial to have
> > > CONFIG_KPROBE_ON_RETHOOK (or somesuch) and just share this.
> >
> > Yes, ideally it should use CONFIG_HAVE_RETHOOK since the rethook arch port
> > must be a copy of the kretprobe implementation. But for safety, I think
> > having CONFIG_KPROBE_ON_RETHOOK is a good idea until replacing all kretprobe
> > implementations.
>
> Masami,
>
> you're respinning this patch to combine
> arch_rethook_trampoline and __kretprobe_trampoline
> right?
Yes, let me send the first patch set (for x86 at first).
BTW, can you review these 2 patches? These are only for the fprobes,
so it can be picked to bpf-next.
https://lore.kernel.org/all/164802091567.1732982.1242854551611267542.stgit@devnote2/T/#u
Thank you,
--
Masami Hiramatsu <[email protected]>