Re: [PATCH bpf-next 06/13] uprobes/x86: Add uprobe syscall to speed up uprobe

[Thomas Weißschuh]

On 2024-12-13 15:51:44+0100, Jiri Olsa wrote:
> On Fri, Dec 13, 2024 at 02:48:00PM +0100, Thomas Weißschuh wrote:
> 
> SNIP
> 
> > > +static int tramp_mremap(const struct vm_special_mapping *sm, struct vm_area_struct *new_vma)
> > > +{
> > > +	return -EPERM;
> > > +}
> > > +
> > > +static struct vm_special_mapping tramp_mapping = {
> > > +	.name   = "[uprobes-trampoline]",
> > > +	.mremap = tramp_mremap,
> > > +};
> > > +
> > > +SYSCALL_DEFINE0(uprobe)
> > > +{
> > > +	struct pt_regs *regs = task_pt_regs(current);
> > > +	struct vm_area_struct *vma;
> > > +	unsigned long bp_vaddr;
> > > +	int err;
> > > +
> > > +	err = copy_from_user(&bp_vaddr, (void __user *)regs->sp + 3*8, sizeof(bp_vaddr));
> > 
> > A #define for the magic values would be nice.
> 
> the 3*8 is to skip 3 values pushed on stack and get the return ip value,
> I'd prefer to keep 3*8 but it's definitely missing explaining comment
> above, wdyt?

A comment sounds good.

> > > +	if (err) {
> > > +		force_sig(SIGILL);
> > > +		return -1;
> > > +	}
> > > +
> > > +	/* Allow execution only from uprobe trampolines. */
> > > +	vma = vma_lookup(current->mm, regs->ip);
> > > +	if (!vma || vma->vm_private_data != (void *) &tramp_mapping) {
> > 
> > vma_is_special_mapping()
> 
> did not know about this function, thanks
> 
> > 
> > > +		force_sig(SIGILL);
> > > +		return -1;
> > > +	}
> > > +
> > > +	handle_syscall_uprobe(regs, bp_vaddr - 5);
> > > +	return 0;
> > > +}
> > > +
> > > +asm (
> > > +	".pushsection .rodata\n"
> > > +	".global uprobe_trampoline_entry\n"
> > > +	"uprobe_trampoline_entry:\n"
> > > +	"endbr64\n"
> > > +	"push %rcx\n"
> > > +	"push %r11\n"
> > > +	"push %rax\n"
> > > +	"movq $" __stringify(__NR_uprobe) ", %rax\n"
> > > +	"syscall\n"
> > > +	"pop %rax\n"
> > > +	"pop %r11\n"
> > > +	"pop %rcx\n"
> > > +	"ret\n"
> > > +	".global uprobe_trampoline_end\n"
> > > +	"uprobe_trampoline_end:\n"
> > > +	".popsection\n"
> > > +);
> > > +
> > > +extern __visible u8 uprobe_trampoline_entry[];
> > > +extern __visible u8 uprobe_trampoline_end[];
> > > +
> > > +const struct vm_special_mapping *arch_uprobe_trampoline_mapping(void)
> > > +{
> > > +	struct pt_regs *regs = task_pt_regs(current);
> > > +
> > > +	return user_64bit_mode(regs) ? &tramp_mapping : NULL;
> > > +}
> > > +
> > > +static int __init arch_uprobes_init(void)
> > > +{
> > > +	unsigned long size = uprobe_trampoline_end - uprobe_trampoline_entry;
> > > +	static struct page *pages[2];
> > > +	struct page *page;
> > > +
> > > +	page = alloc_page(GFP_HIGHUSER);
> > 
> > That page could be in static memory, removing the need for the explicit
> > allocation. It could also be __ro_after_init.
> > Then tramp_mapping itself can be const.
> 
> hum, how would that look like? I think that to get proper page object
> you have to call alloc_page or some other page alloc family function..
> what do I miss?

static u8 trampoline_page[PAGE_SIZE] __ro_after_init __aligned(PAGE_SIZE);
static struct page *tramp_mapping_pages[2] __ro_after_init;

static const struct vm_special_mapping tramp_mapping = {
	.name   = "[uprobes-trampoline]",
	.pages  = tramp_mapping_pages,
	.mremap = tramp_mremap,
};

static int __init arch_uprobes_init(void)
{
	...
	trampoline_pages[0] = virt_to_page(trampoline_page);
	...
}

Untested, but it's similar to the stuff the vDSO implementations are
doing which I am working with at the moment.

> > 
> > Also this seems to waste the page on 32bit kernels.
> 
> it's inside CONFIG_X86_64 ifdef
> 
> > 
> > > +	if (!page)
> > > +		return -ENOMEM;
> > > +	pages[0] = page;
> > > +	tramp_mapping.pages = (struct page **) &pages;
> > 
> > tramp_mapping.pages = pages; ?
> 
> I think the compiler will cry about *pages[2] vs **pages types mismatch,
> but I'll double check that

It compiles for me.

> thanks,
> jirka
> 
> > 
> > > +	arch_uprobe_copy_ixol(page, 0, uprobe_trampoline_entry, size);
> > > +	return 0;
> > > +}
> > > +
> > > +late_initcall(arch_uprobes_init);
> > > +
> > >  /*
> > >   * If arch_uprobe->insn doesn't use rip-relative addressing, return
> > >   * immediately.  Otherwise, rewrite the instruction so that it accesses
> > 
> > [..]