On Fri, Jan 21, 2011 at 10:41:54PM +0100, matthieu castet wrote:Konrad Rzeszutek Wilk a écrit :Ok,- * .data and .bss should always be writable.Hold it.. ccache is a wonderful tool but I think I've just "rebuilt" the
+ * .data and .bss should always be writable, but xen won't like
+ * if we make page table rw (that live in .data or .bss)
*/
+#ifdef CONFIG_X86_32
if (within(address, (unsigned long)_sdata, (unsigned long)_edata) ||
- within(address, (unsigned long)__bss_start, (unsigned long)__bss_stop))
- pgprot_val(required) |= _PAGE_RW;
+ within(address, (unsigned long)__bss_start, (unsigned long)__bss_stop)) {
+ unsigned int level;
+ if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
+ pgprot_val(forbidden) |= _PAGE_RW;
+ }
+#endif
#if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA)
fyi, it does make it boot.
binaries with the .bss HPAGE_ALIGN aligment by mistake, so this path got never
taken.
ATM I saw the following solution to solve the problem :
1) remove the data/bss check in static_protections, it was introduced by NX patches (64edc8ed). But I am not sure it
is really needed anymore.
2) add ". = ALIGN(HPAGE_SIZE)" somewhere after init section. But if we want not to be allocated in image we
should put it before bss. And if we want to be freed after init, we should put before .init.end.
This mean moving .smp_locks (and .data_nosave when x86 will be added) before init section. I have no idea of the impact.
3) add some logic in arch/x86/xen/mmu.c, that will ignore RW page setting for the page table marked RO.
4) make static_protections take and old_prot argument, and only apply RW .data/.bss requirement if page is already RW.
If possible I will go for 1).
Sounds good. Just send me the patch and I will test it.