[Xen-devel] Explanation of set_pte_at

[Michael Abd-El-Malek <mabdelmalek@xxxxxxx>]

Hi,

Can someone please explain the function set_pte_at to me?  In  
particular, I'm trying to understand the page fault handling in a PV  
64-bit Linux.
Here's the sequence that I trace:
1) The Linux page fault handler at linux-2.6.18-xen.hg/mm/ 
memory.c:do_no_page runs.
2) After getting the page in memory, it calls set_pte_at.
3) set_pte_at has the following definition:

#define set_pte_at(_mm,addr,ptep,pteval) do {
       if (((_mm) != current->mm && (_mm) != &init_mm) ||          \
     HYPERVISOR_update_va_mapping((addr), (pteval), 0))          \
        set_pte((ptep), (pteval));                                  \

In the case of a page fault, I think _mm will be equal to init_mm.   
As a result, the first if condition fails, and we end up calling  
HYPERVISOR_update_va_mapping.   If that succeeds, then we're done.  
Otherwise, we call set_pte which just sets the PTE directly.  Since in  
PV Linux, page tables don't have writeable mappings, this write causes  
an instruction fault and Xen's ptwr_emulated_update tries to perform  
the page table update.
So here are my questions:
1) What is the purpose of the first if condition ((_mm) != current->mm  
&& (_mm) != &init_mm)?  If the fault happens in a kernel virtual  
memory context, why do we always want to call set_pte?  Why don't we  
just call HYPERVISOR_update_va_mapping?  Does ptwr_emulated_update  
handle a larger subset of the page table updates logic than  
HYPERVISOR_update_va_mapping?
2) On a related note, if we do call HYPERVISOR_update_va_mapping and  
it fails, why do we retry the page table update through  
ptwr_emulated_update?
Thanks!
Mike
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel