Hi Kevin,
Thanks for your explanation.
Sorry, I'd like you to explain this once again. Please look at the
below figure.
1) Instruction TLB Fault ---+
|
+----------------------+
|
+---> ENTRY(iltb_miss)
/* Check ifa (It was VHPT_CCHAIN_LOOKUP before here) */
mov r16 = cr.ifa
extr.u r17=r16,59,5
cmp.eq p6,p0=0x1e,r17
(p6) br.cond.spnt late_alt_itlb_miss -----+
cmp.eq p6,p0=0x1d,r17 |
(p6) br.cond.spnt late_alt_itlb_miss ---+ |
| |
| |
2) Alternate Instruction TLB Fault ---+ | |
| | |
+--------------------------------+ | |
| | |
+---> ENTRY(alt_itlb_miss) | |
mov r16=cr.ifa | |
| |
late_alt_itlb_miss: <-------------------+-+
/* Check cpl */
cmp.ne p8,p0=r0,r23
or r19=r17,r19
or r19=r19,r18
(p8) br.cond.spnt page_fault
+ /* Check ifa with my patch */
+ extr.u r22=r16,59,5
+ cmp.ne p8,p0=0x1e,r22
+ (p8) br.cond.spnt 1f ----------+
|
itc.i r19 |
mov pr=r31,-1 |
rfi |
|
+ 1: <---------------------------+
+ FORCE_CRASH
If case 1), I think that a FORCE_CRASH and ifa checking is unnecessary
according to your explanation.
If case 2), I think that a FORCE_CRASH and ifa checking is necessary.
Because, I thought that Xen may use a wrong address.
If case 2), does Xen trust only cpl?
Best regards,
Kan
Tian, Kevin wrote:
>>From: Masaki Kanno [mailto:kanno.masaki@xxxxxxxxxxxxxx]
>>Sent: 2006定4埖21晩 18:56
>>>>
>>>>Hi Kan,
>>>>
>>>> Thanks, this looks like exactly what we need. If there are no
>>other
>>>>comments, please send me this patch w/ a Signed-off-by and we can
>>get
>>>>it
>>>>in tree. BTW, glad to hear you're working on the FPSWA issue and
>>are
>>>>making good progress! Thanks,
>>>>
>>>> Alex
>>>
>>>Seems OK. One small comment is that we may also remove
>>>FORCE_CRASH completely since the assumption to add that
>>>check doesn't exist now. Actually VHPT_CCHAIN_LOOKUP
>>>already makes check upon VMM area to decide whether jumping
>>>to alt_itlb_miss handler. In this case, simply removing
>>>FORCE_CRASH line can also work. :-)
>>
>>If alt_itlb_fault occurred, we need ifa checking and FORCE_CRASH,
>>don't we?
>>Therefore I don't need to change my patch, do I?
>>
>
>The check is already made before jumping to alt_itlb_miss.
>Also architecturally there's no limitation to prevent uncacheable
>instruction falling into that category. So I think there's no need
>for existence of FORCE_CRASH there, right? :-)
>
>Thanks,
>Kevin
>
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|