On Fri, Jul 25, 2008 at 05:47:37PM +0900, SUZUKI Kazuhiro wrote:
The following patch fixes the mca handler so as not to destroy ar
and some bugs.
Thank you for fixing some bugs and it looks basically good.
Some comments below.
@@ -524,24 +457,111 @@ ia64_reload_tr:
srlz.d
;;
#ifdef XEN
-.reload_vhpt:
- // 5. VHPT
- GET_THIS_PADDR(r1, inserted_vhpt);;
- cmp.eq p7,p0=r2,r0
-(p7) br.cond.sptk .overlap_vhpt // vhpt isn't mapped.
+ // 5. shared_info
+ GET_THIS_PADDR(r2, inserted_shared_info);;
+ ld8 r16=[r2]
+ mov r18=XSI_SHIFT<<2
+ movl r20=__pgprot(__DIRTY_BITS | _PAGE_PL_PRIV | _PAGE_AR_RW)
+ ;;
+ GET_THIS_PADDR(r2, domain_shared_info);;
+ ld8 r17=[r2]
+ ;;
+ dep r17=0,r17,60,4
+ ;;
+ or r17=r17,r20 // construct PA | page properties
+ mov cr.itir=r18
+ mov cr.ifa=r16
+ ;;
+ mov r16=IA64_TR_SHARED_INFO
+ ;;
+ itr.d dtr[r16]=r17 // wire in new mapping...
+ ;;
+ srlz.d
+ ;;
Unconditionally pinning down shared_info into inserted_shared_info
is wrong because shared_info is shared only with PV domain and xen VMM.
So In VMX domain case, it shouldn't pinned down. Otherwise VMX guest
wrongly accesses to shared_info.
In ia64_do_tlb_purge() case, unconditionally purging
DTR[IA64_TR_SHARED_INFO] is okay, but unconditionally inserting
DTR[IA64_TR_SHARED_INFO] is bad.
thanks,
15 messages in com.xensource.lists.xen-ia64-develRe: [PATCH] Fix mca handler so as not...
.patch