AArch64: SCTLR_EL1.BT0 set incorrectly in user mode

Host environment

• Operating system: Ubuntu 20.04
• OS/kernel version: 5.4.0-84-generic
• Architecture: ARM
• QEMU flavor: qemu-aarch64
• QEMU version: v7.0.0-rc4
• QEMU command line:

  ../qemu/build/qemu-aarch64 --cpu max -L ~/fedora-rootfs/fedora_aarch64 -g 4022 ./sample-big hi

Emulated/Virtualized environment

• Operating system: Fedora 34
• OS/kernel version: n/a
• Architecture: ARM

Description of problem

PACIASP normally acts as a BTI landing pad, but not in every situation. When SCTLR_EL1.BT is set, PACIASP checks that the indirect branch originates from X16 or X17 when the indirect branch is taken from a BTI guarded area. Linux sets this bit, ideally QEMU-user should too. This sample program should crash with a SIGILL if QEMU is working correctly, otherwise it will crash with a SIGSEGV.

#include <stdint.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <stdio.h>
#include <sys/mman.h>

// PACIASP is a valid BTI landing pad, but there are some conditions
// under Linux which sets SCTLR_ELx.BT0 = 1. In this mode, a branch
// onto a PACIASP landing pad is only valid if it originates from
// x16 or x17 (i.e. br x17 is OK, br x3 is not).
// More info on page D5-4851 of the Arm Architecture Reference Manual (ARM DDI 0487H.a).

// Sample function which starts with a paciasp instruction
// (comes from -mbranch-protection=pac-ret+leaf)
void indirect_fn(int i) {
    // paciasp instruction inserted here - should crash with SIGILL here if everything's operating OK.
    i = i+1;
    // Can't access this function from the copied location, so will segfault.
    fprintf(stderr, "reachable\n");
}

int main(int argc, char **argv) {
    // It's difficult to get a whole binary BTI compatible without the appropriate crtbegin etc
    // so instead map a page and copy the sample function there.
    void *e = mmap(0, getpagesize(), PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
    if (e == MAP_FAILED) {
        return 1;
    }
    memcpy(e, (void*)indirect_fn, 64);
    mprotect(e, getpagesize(), PROT_READ | PROT_EXEC | PROT_BTI);

    // paciasp is a valid landing pad if the branch comes from an unprotected area,
    // so to ensure that we're protected - assemble an intermediate shim that's also PROT_BTI.
    void *f = mmap(0, getpagesize(), PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
    if (f == MAP_FAILED) {
        return 1;
    }
    uint32_t *x = (uint32_t*)f;
    x[0] = 0xd503245fuL; // bti c
    x[1] = 0xd61f0060uL; // br x1 - n.b. must be BR
    mprotect(f, getpagesize(), PROT_READ | PROT_EXEC | PROT_BTI);

    // Jump to the shim
    asm volatile (
        "mov x3, %0\n"
        "mov x2, %1\n"
        "blr x2\n"
    : : "p"(e), "p"(f) : "x2", "x3");

    // Execution should not reach here
    return 1;
}

Steps to reproduce

1. Compile with clang-12 -g --sysroot=/work/home/fedora-rootfs/fedora_aarch64 -o sample --target=aarch64-linux-gnu -mbranch-protection=pac-ret+leaf -march=armv8-a -O1 -g sample.c or similar.
2. Run with ../qemu/build/qemu-aarch64 --cpu max -L ~/fedora-rootfs/fedora_aarch64 sample

Additional information

n/a