Re: [RFC PATCH v4 2/4] x86/sgx: Implement support for MADV_WILLNEED

["Huang, Kai" <kai.huang@intel.com>]

On Tue, 2023-02-14 at 13:18 -0600, Haitao Huang wrote:
> Hi Kai
> 
> On Tue, 14 Feb 2023 03:47:24 -0600, Huang, Kai <kai.huang@intel.com> wrote:
> 
> > On Fri, 2023-01-27 at 20:55 -0800, Haitao Huang wrote:
> > > @@ -97,10 +99,81 @@ static int sgx_mmap(struct file *file, struct  
> > > vm_area_struct *vma)
> > >  	vma->vm_ops = &sgx_vm_ops;
> > >  	vma->vm_flags |= VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP | VM_IO;
> > >  	vma->vm_private_data = encl;
> > > +	vma->vm_pgoff = PFN_DOWN(vma->vm_start - encl->base);
> > >   	return 0;
> > >  }
> > 
> > Perhaps I am missing something, but above change looks weird.  
> > Conceptually, it doesn't/shouldn't belong to this series, which  
> > essentially
> > preallocates and does EAUG EPC pages for a (or part of) given enclave.   
> > The EAUG
> > logic should already be working for the normal fault path, which means  
> > the code
> > change above either: 1) has been done at other place; 2) isn't needed.
> > 
> > I have kinda forgotten the userspace sequence to create an enclave.  If  
> > I recall
> > correctly, you do below to create an enclave:
> > 
> > 	1) encl_fd = open("/dev/sgx_enclave");
> > 	2) encl_addr = mmap(encl_fd, encl_size, 0 /* pgoff */);
> > 	3) IOCTL(ECREATE, encl_addr, encl_size);
> > 
> > Would the above code change break the "mmap()" in above step 2?
> > 	
> 
> No, vm_pgoff was not used previously for enclave VMAs. I had to add this  
> because the offset passed to sgx_fadvise is relative to file base and  
> calculated in mm/madvise.c like this:
> 
>          offset = (loff_t)(start - vma->vm_start)
>                          + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);

But shouldn't 'offset is relative to the file base' be conceptually correct from
the fadvice()'s point of view?

I think you should do:

	encl_offset = offset + encl->base;

inside sgx_fadvice()?

> 
> I had a comment in first version but removed it based on Jarkko's  
> suggestion here: https://lore.kernel.org/all/Y2B0jBsG6HE4KVk7@kernel.org/
> 
> The original comments probably seemed redundant to the definitions of the  
> vm_pgoff field and the fadvise interface. But let me know if we need add a  
> more helpful version of comments or any suggestion on the comments.

I still think this code change is wrong.

For instance, IIUC, it at least breaks the case where enclave hasn't been
created/initialized, where encl->base == 0 (although normal code path doesn't
use vm_pgoff, conceptually it's still wrong IIUC).

Maybe I am missing something?