Possible regression in QEMU 5.0.0 after CVE-2020-10702 (segmentation fault)

[Héctor Molinero Fernández]

I've come across a very specific situation, but I'm sure it could be replicated in other cases.

In QEMU 5.0.0 when I use user emulation with a cURL binary for aarch64 and connect to a server using TLS 1.2 and ECDHE-ECDSA-CHACHA20-POLY1305 cypher a segmentation fault occurs.

I attach a Dockerfile that reproduces this crash and the strace output with and without the de0b1bae6461f67243282555475f88b2384a1eb9 commit reverted.

[Richard Henderson (rth)]

This is a compiler bug affecting (at least) libcrypto.so.1.1:

  179d90: d503233f paciasp
  179d94: a9bb7bfd stp x29, x30, [sp, #-80]!
...
  17a400: d50323bf autiasp
  17a404: f84507fd ldr x29, [sp], #80
  17a408: d65f03c0 ret

The PAC happens with the initial sp:

  X30=0000005501de55fc SP=00000055018477a0

while the AUTH happens with the decremented sp:

  X30=0011005501de55fc SP=0000005501847750

Since the salt (sp) is different for the two operations, the
authorization should and does fail:

  X30=0020005501de55fc

Note bit 53 is now set in x30, which is the error indication.

The compiler must move the authiasp down below the ldr pop.