Hi Sean,
On Wed, May 12, 2021 at 05:31:03PM +0000, Sean Christopherson wrote:
This got me looking at the flows that "inject" #PF, and I'm pretty sure there are bugs in __vc_decode_user_insn() + insn_get_effective_ip().
Problem #1: __vc_decode_user_insn() assumes a #PF if insn_fetch_from_user_inatomic() fails, but the majority of failure cases in insn_get_seg_base() are #GPs, not #PF.
res = insn_fetch_from_user_inatomic(ctxt->regs, buffer); if (!res) { ctxt->fi.vector = X86_TRAP_PF; ctxt->fi.error_code = X86_PF_INSTR | X86_PF_USER; ctxt->fi.cr2 = ctxt->regs->ip; return ES_EXCEPTION; }
Problem #2: Using '0' as an error code means a legitimate effective IP of '0' will be misinterpreted as a failure. Practically speaking, I highly doubt anyone will ever actually run code at address 0, but it's technically possible. The most robust approach would be to pass a pointer to @ip and return an actual error code. Using a non-canonical magic value might also work, but that could run afoul of future shenanigans like LAM.
ip = insn_get_effective_ip(regs); if (!ip) return 0;
Your observations are all correct. I put some changes onto this patch-set to fix these problems.
Regards,
Joerg