Re: [PATCH 3/3] x86/64/kexec: Rewrite init_transition_pgtable() with kernel_ident_mapping_init()

[Huang, Kai]

On Thu, 2024-07-04 at 16:44 +0300, kirill.shutemov@xxxxxxxxxxxxxxx wrote:
> On Wed, Jul 03, 2024 at 11:06:21AM +0000, Huang, Kai wrote:
> > >  static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
> > >  {
> > > -	pgprot_t prot = PAGE_KERNEL_EXEC_NOENC;
> > > -	unsigned long vaddr, paddr;
> > > -	int result = -ENOMEM;
> > > -	p4d_t *p4d;
> > > -	pud_t *pud;
> > > -	pmd_t *pmd;
> > > -	pte_t *pte;
> > > +	struct x86_mapping_info info = {
> > > +		.alloc_pgt_page	= alloc_transition_pgt_page,
> > > +		.context	= image,
> > > +		.page_flag	= __PAGE_KERNEL_LARGE_EXEC,
> > > +		.kernpg_flag	= _KERNPG_TABLE_NOENC,
> > > +		.offset = __START_KERNEL_map - phys_base,
> > > +	};
> > > +	unsigned long mstart = PAGE_ALIGN_DOWN(__pa(relocate_kernel));
> > > +	unsigned long mend = mstart + PAGE_SIZE;
> > >  
> > > -	vaddr = (unsigned long)relocate_kernel;
> > > -	paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
> > 
> > Perhaps I am missing something, but this seems a functional change to me.
> > 
> > IIUC the page after image->control_code_page is allocated when loading the
> > kexec kernel image.  It is a different page from the page where the
> > relocate_kernel code resides in.
> > 
> > The old code maps relocate_kernel kernel VA to the page after the
> > control_code_page.  Later in machine_kexec(), the relocate_kernel code is
> > copied to that page so the mapping can work for that:
> > 
> > 	control_page = page_address(image->control_code_page) + PAGE_SIZE;
> > 	__memcpy(control_page, relocate_kernel,
> > KEXEC_CONTROL_CODE_MAX_SIZE);
> > 
> > The new code in this patch, however, seems just maps the relocate_kernel VA
> > to the PA of the relocate_kernel, which should be different from the old
> > mapping.
> 
> Yes, original code maps at relocate_kernel() VA the page with copy of the
> relocate_kernel() in control_code_page. But it is safe to map original
> relocate_kernel() page there as well as it is not going to be overwritten
> until swap_pages(). We are not going to use original relocate_kernel()
> page after RET at the end of relocate_kernel().

I am not super familiar with this, but this doesn't seem 100% safe to me.

E.g, did you consider the kexec jump case?

The second half of control page is also used to store registers in kexec
jump.  If the relocate_kernel VA isn't mapped to the control page, then IIUC
after jumping back to old kernel it seems we won't be able to read those
registers back?

> 
> Does it make any sense?
> 
> I will try to explain it in the commit message in the next version.
> 

I think even it's safe to change to map to the relocate_kernel() page, it
should be done in a separate patch.  This patch should just focus on removing
the duplicated page table setup code.