[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Xen-ia64-devel] alt_itlb_miss?
Thank for your explanation.
Tian, Kevin wrote:
> Thanks for detail figure. From architecture correctness, I think
>your patch is yes required though case 2 is only walked before dom0
>loads rr7 (After that, vhpt table will be enabled from then on and case
>1 is the only path then).
I misunderstood about VHPT. When Xen code is running, I thought that
VHPT was always disabled. Then, I also think that I have only to remove
FORCE_CRASH from alt_itlb_miss.
> So please re-submit the patch again and it's better if you could
>move the check to the point before late_alt_itlb_miss to avoid
>duplicated check in case 2. Also, it's better to jump to page_fault
>instead of crash the whole xen here which is over-killed. :-)
>>From: Masaki Kanno [mailto:kanno.masaki@xxxxxxxxxxxxxx]
>>Sent: 2006定4埖24晩 15:40
>>To: Tian, Kevin
>>Cc: Alex Williamson; Isaku Yamahata; xen-ia64-devel
>>Subject: Re: [Xen-ia64-devel] alt_itlb_miss?
>>Thanks for your explanation.
>>Sorry, I'd like you to explain this once again. Please look at the
>>1) Instruction TLB Fault ---+
>> +---> ENTRY(iltb_miss)
>> /* Check ifa (It was VHPT_CCHAIN_LOOKUP before
>> 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
>>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?
>>Tian, Kevin wrote:
>>>>From: Masaki Kanno [mailto:kanno.masaki@xxxxxxxxxxxxxx]
>>>>Sent: 2006協4・Q1賦 18:56
>>>>>> Thanks, this looks like exactly what we need. If there are no
>>>>>>comments, please send me this patch w/ a Signed-off-by and we
>>>>>>in tree. BTW, glad to hear you're working on the FPSWA issue
>>>>>>making good progress! Thanks,
>>>>>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,
>>>>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? :-)
Xen-ia64-devel mailing list