Re: [PATCH RFCv2 12/18] uprobes/x86: Add support to optimize uprobes
From: Jiri Olsa <hidden>
Date: 2025-02-28 23:00:51
Also in:
bpf, lkml
On Mon, Feb 24, 2025 at 03:01:44PM +0100, Jiri Olsa wrote: SNIP
quoted hunk ↗ jump to hunk
@@ -1523,15 +1698,23 @@ arch_uretprobe_hijack_return_addr(unsigned long trampoline_vaddr, struct pt_regs { int rasize = sizeof_long(regs), nleft; unsigned long orig_ret_vaddr = 0; /* clear high bits for 32-bit apps */ + unsigned long off = 0; + + /* + * Optimized uprobe goes through uprobe trampoline which adds 4 8-byte + * values on stack, check uprobe_trampoline_entry for details. + */ + if (!swbp) + off = 4*8;
ok, now when I started to add the missing register modifications in uprobe syscall, I realized we will modify the regs->sp appropriately already in the uprobe syscall (before the code above is hit) so we don't need this code and we can get rid of the swbp flag and patch#7 completely jirka
- if (copy_from_user(&orig_ret_vaddr, (void __user *)regs->sp, rasize))
+ if (copy_from_user(&orig_ret_vaddr, (void __user *)regs->sp + off, rasize))
return -1;
/* check whether address has been already hijacked */
if (orig_ret_vaddr == trampoline_vaddr)
return orig_ret_vaddr;
- nleft = copy_to_user((void __user *)regs->sp, &trampoline_vaddr, rasize);
+ nleft = copy_to_user((void __user *)regs->sp + off, &trampoline_vaddr, rasize);
if (likely(!nleft)) {
if (shstk_update_last_frame(trampoline_vaddr)) {
force_sig(SIGSEGV);SNIP