Don't permit mappings of invalid physical addresses on x86 via /dev/mem.

[jhb]

Summary

Don't permit mappings of invalid physical addresses on x86 via /dev/mem.

While here, add a bounds check on the physical address used with
memrw on i386. (amd64 already has a check)

Test Plan

• mmap an invalid address via /dev/mem. Triggers a panic in an unpatched kernel on access.
• I have only tested amd64 so far, need to verify on i386 still.

[kib]

I think that this may be not completely correct for i386. cpu_maxphyaddr contains raw value from CPUID. On non-PAE systems the result could be that offsets are allowed behind 4G. The result would be non-fatal, but incorrect.

x86/nexus.c has same bug, it seems.

[jhb]

We always set cpu_maxphyaddr. If the cpuid leaf isn't available we set it to either 32 or 36 depending on PAE:

if (cpu_exthigh >= 0x80000008) {
        do_cpuid(0x80000008, regs);
        cpu_maxphyaddr = regs[0] & 0xff;
        cpu_procinfo2 = regs[2];
} else {
        cpu_maxphyaddr = (cpu_feature & CPUID_PAE) != 0 ? 36 : 32;
}

[kib]

In D7408#154024, @jhb wrote:

We always set cpu_maxphyaddr. If the cpuid leaf isn't available we set it to either 32 or 36 depending on PAE:

I mean that if CPUID leaf is available and typically wider than 32, on !PAE config the offsets > 4G are not representable. They will be wrapped mod 4G, which, as I said above, not fatal, but not correct either.

[jhb]

Hmm, so perhaps the code to set cpu_maxphyaddr should force it to 32 in the i386 && !PAE case?

[kib]

In D7408#154070, @jhb wrote:

Hmm, so perhaps the code to set cpu_maxphyaddr should force it to 32 in the i386 && !PAE case?

I suspect this might be too limiting overall. If you want to avoid duplicated limiting code repeated several times for i386, perhaps inline function is due in x86_var.h ?

static inline vm_paddr_t
cpu_getmaxusedphyaddr(void)
{
#if defined(__i386__) && !defined(PAE)
   return (0xffffffffU);
#else
   return (1ULL << cpu_maxphyaddr);
#endif
}

I do not think that PAE_TABLES option should result in the full width return.

[jhb]

• Export maxphyaddr() as x86-wide cpu_getmaxphyaddr().
• Drop debug printf.
• Rebase after EFER_NXE fix and /dev/kmem mmap removal.

[jhb]

sys/i386/i386/mem.c

111 ↗

(On Diff #21706)

It was important to check uio_offset and not 'pa' to reject addresses above 4G on i386 'pa' was only 32 bits and would silently truncate on assignment (I originally checked pa). Also, using trunc_page() instead of inlining it seemed more consistent (I can't recall if it mattered for PAE as well)

[jhb]

Oh, and I (finally) tested this on both PAE and !PAE on i386 (PAE testing was stuck on the EFER_NXE bug)