On Sat, Jul 13, 2019 at 08:07:52PM +0300, Jarkko Sakkinen wrote:
> +static unsigned long sgx_get_unmapped_area(struct file *file,
> + unsigned long addr,
> + unsigned long len,
> + unsigned long pgoff,
> + unsigned long flags)
> +{
> + if (flags & MAP_PRIVATE)
> + return -EINVAL;
> +
> + if (flags & MAP_FIXED)
> + return addr;
> +
> + if (len < 2 * PAGE_SIZE || len & (len - 1))
> + return -EINVAL;
> +
> + addr = current->mm->get_unmapped_area(file, addr, 2 * len, pgoff,
> + flags);
> + if (IS_ERR_VALUE(addr))
> + return addr;
> +
> + addr = (addr + (len - 1)) & ~(len - 1);
> +
> + return addr;
> +}
Thinking about this more, I don't think the driver should verify or adjust
@addr and @len during non-fixed mmap(). There is no requirement that
userspace must map the full ELRANGE, or that an enclave's ELRANGE can
*never* be mapped to non-EPC. The architectural requirement is that an
access that hits ELRANGE must map to the EPC *if* the CPU is executing the
associated enclave.
This means that a process can have random mappings that overlap an
enclave's ELRANGE while the enclave is active, it just can't access that
memory from within the enclave.
A good example is a large enclave, e.g. 3gb of actual code/data, that
userspace wants to locate below 4gb for some reason, e.g. running an
unmodified non-enclave binary under graphene. Requiring @addr+@len to be
naturally sized and aligned will force @addr to 0x0, which is often
disallowed by the kernel.
