Re: [PATCH bpf-next 05/13] uprobes: Add mapping for optimized uprobe trampolines

[Jiri Olsa <olsajiri@gmail.com>]

On Thu, Dec 12, 2024 at 05:01:52PM -0800, Andrii Nakryiko wrote:

SNIP

> > ---
> >  include/linux/uprobes.h |  12 +++++
> >  kernel/events/uprobes.c | 114 ++++++++++++++++++++++++++++++++++++++++
> >  kernel/fork.c           |   1 +
> >  3 files changed, 127 insertions(+)
> >
> 
> Ran out of time for today, will continue tomorrow for the rest of
> patches. Some comments below.

thanks!

> 
> The numbers are really encouraging, though!
> 
> > diff --git a/include/linux/uprobes.h b/include/linux/uprobes.h
> > index 8843b7f99ed0..c4ee755ca2a1 100644
> > --- a/include/linux/uprobes.h
> > +++ b/include/linux/uprobes.h
> > @@ -16,6 +16,7 @@
> >  #include <linux/types.h>
> >  #include <linux/wait.h>
> >  #include <linux/timer.h>
> > +#include <linux/mutex.h>
> >
> >  struct uprobe;
> >  struct vm_area_struct;
> > @@ -172,6 +173,13 @@ struct xol_area;
> >
> >  struct uprobes_state {
> >         struct xol_area         *xol_area;
> > +       struct hlist_head       tramp_head;
> > +};
> > +
> 
> should we make uprobe_state be linked by a pointer from mm_struct
> instead of increasing mm for each added field? right now it's
> embedded, I don't think it's problematic to allocate it on demand and
> keep it until mm_struct is freed

seems like good idea, I'll check on that

> 
> > +struct uprobe_trampoline {
> > +       struct hlist_node       node;
> > +       unsigned long           vaddr;
> > +       atomic64_t              ref;
> >  };
> >
> >  extern void __init uprobes_init(void);
> > @@ -220,6 +228,10 @@ extern int arch_uprobe_verify_opcode(struct arch_uprobe *auprobe, struct page *p
> >                                      unsigned long vaddr, uprobe_opcode_t *new_opcode,
> >                                      int nbytes);
> >  extern bool arch_uprobe_is_register(uprobe_opcode_t *insn, int nbytes);
> > +extern struct uprobe_trampoline *uprobe_trampoline_get(unsigned long vaddr);
> > +extern void uprobe_trampoline_put(struct uprobe_trampoline *area);
> > +extern bool arch_uprobe_is_callable(unsigned long vtramp, unsigned long vaddr);
> > +extern const struct vm_special_mapping *arch_uprobe_trampoline_mapping(void);
> >  #else /* !CONFIG_UPROBES */
> >  struct uprobes_state {
> >  };
> > diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
> > index 8068f91de9e3..f57918c624da 100644
> > --- a/kernel/events/uprobes.c
> > +++ b/kernel/events/uprobes.c
> > @@ -615,6 +615,118 @@ set_orig_insn(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long v
> >                         (uprobe_opcode_t *)&auprobe->insn, UPROBE_SWBP_INSN_SIZE);
> >  }
> >
> > +bool __weak arch_uprobe_is_callable(unsigned long vtramp, unsigned long vaddr)
> 
> bikeshedding some more, I still find "is_callable" confusing. How
> about "is_reachable_by_call"? slightly verbose, but probably more
> meaningful?

yep, more precise, will change

> 
> > +{
> > +       return false;
> > +}
> > +
> > +const struct vm_special_mapping * __weak arch_uprobe_trampoline_mapping(void)
> > +{
> > +       return NULL;
> > +}
> > +
> > +static unsigned long find_nearest_page(unsigned long vaddr)
> > +{
> > +       struct mm_struct *mm = current->mm;
> > +       struct vm_area_struct *vma, *prev;
> > +       VMA_ITERATOR(vmi, mm, 0);
> > +
> > +       prev = vma_next(&vmi);
> 
> minor: we are missing an opportunity to add something between
> [PAGE_SIZE, <first_vma_start>). Probably fine, but why not?

true, will add that check

> 
> > +       vma = vma_next(&vmi);
> > +       while (vma) {
> > +               if (vma->vm_start - prev->vm_end  >= PAGE_SIZE) {
> > +                       if (arch_uprobe_is_callable(prev->vm_end, vaddr))
> > +                               return prev->vm_end;
> > +                       if (arch_uprobe_is_callable(vma->vm_start - PAGE_SIZE, vaddr))
> > +                               return vma->vm_start - PAGE_SIZE;
> > +               }
> > +
> > +               prev = vma;
> > +               vma = vma_next(&vmi);
> > +       }
> > +
> > +       return 0;
> > +}
> > +
> 
> [...]
> 
> > +struct uprobe_trampoline *uprobe_trampoline_get(unsigned long vaddr)
> > +{
> > +       struct uprobes_state *state = &current->mm->uprobes_state;
> > +       struct uprobe_trampoline *tramp = NULL;
> > +
> > +       hlist_for_each_entry(tramp, &state->tramp_head, node) {
> > +               if (arch_uprobe_is_callable(tramp->vaddr, vaddr)) {
> > +                       atomic64_inc(&tramp->ref);
> > +                       return tramp;
> > +               }
> > +       }
> > +
> > +       tramp = create_uprobe_trampoline(vaddr);
> > +       if (!tramp)
> > +               return NULL;
> > +
> > +       hlist_add_head(&tramp->node, &state->tramp_head);
> > +       return tramp;
> > +}
> > +
> > +static void destroy_uprobe_trampoline(struct uprobe_trampoline *tramp)
> > +{
> > +       hlist_del(&tramp->node);
> > +       kfree(tramp);
> 
> hmm... shouldn't this be RCU-delayed (RCU Tasks Trace for uprobes),
> otherwise we might have some CPU executing code in that trampoline,
> no?

so we call destroy_uprobe_trampoline in 2 scenarios:

  - from uprobe_trampoline_put (in __arch_uprobe_optimize) when we failed
    to optimize the uprobe, so no task can execute it at that point

  - from clear_tramp_head as part of the uprobe trampolines cleanup
    (__mmput -> uprobe_clear_state) at which point the task should be dead

jirka

> 
> > +}
> > +
> 
> [...]