diff mbox

[Xen-devel,v2,5/6] xen/arm: traps: Avoid unnecessary VA -> IPA translation in abort handlers

Message ID 1469639378-9244-6-git-send-email-julien.grall@arm.com
State Superseded
Headers show

Commit Message

Julien Grall July 27, 2016, 5:09 p.m. UTC
Translating a VA to a IPA is expensive. Currently, Xen is assuming that
HPFAR_EL2 is only valid when the stage-2 data/instruction abort happened
during a translation table walk of a first stage translation (i.e S1PTW
is set).

However, based on the ARM ARM (D7.2.34 in DDI 0487A.j), the register is
also valid when the data/instruction abort occured for a translation
fault.

With this change, the VA -> IPA translation will only happen for
permission faults that are not related to a translation table of a
first stage translation.

Signed-off-by: Julien Grall <julien.grall@arm.com>

---
    Changes in v2:
        - Use fsc in the switch in do_trap_data_abort_guest
---
 xen/arch/arm/traps.c | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

Comments

Julien Grall July 27, 2016, 5:40 p.m. UTC | #1
On 27/07/16 18:28, Sergej Proskurin wrote:
> Hi Julien,

Hello Sergej,

Please reply to all (or at least CC me) when answering to xen-devel. 
Otherwise your mail may be missed or the answer may be delayed because I 
filter any mail where I am not the recipients (a lot of people do this 
on the ML).

>
> On 07/27/2016 07:09 PM, Julien Grall wrote:
>> @@ -2442,7 +2458,7 @@ static void do_trap_data_abort_guest(struct cpu_user_regs *regs,
>>      info.gva = READ_SYSREG64(FAR_EL2);
>>  #endif
>>
>> -    if ( dabt.s1ptw )
>> +    if ( hpfar_is_valid(hsr.iabt.s1ptw, fsc) )
>
> I belive this should be:
>
> hpfar_is_valid(hsr.*dabt*.s1ptw, fsc)

You are right, I did not spot it during my testing because the field is 
exactly the same bits in the HSR.

>
>>          info.gpa = get_faulting_ipa(info.gva);
>>      else
>>      {
>> @@ -2451,7 +2467,7 @@ static void do_trap_data_abort_guest(struct cpu_user_regs *regs,
>>              return; /* Try again */
>>      }
>>
>> -    switch ( dabt.dfsc & ~FSC_LL_MASK )
>> +    switch ( fsc )
>>      {
>>      case FSC_FLT_PERM:
>>      {

Regards,
Julien Grall Aug. 17, 2016, 11:11 a.m. UTC | #2
On 17/08/16 03:19, Shanker Donthineni wrote:
> Hi Julien,

Hello Shanker,

> On 07/27/2016 12:09 PM, Julien Grall wrote:
>> Translating a VA to a IPA is expensive. Currently, Xen is assuming that
>> HPFAR_EL2 is only valid when the stage-2 data/instruction abort happened
>> during a translation table walk of a first stage translation (i.e S1PTW
>> is set).
>>
>> However, based on the ARM ARM (D7.2.34 in DDI 0487A.j), the register is
>> also valid when the data/instruction abort occured for a translation
>> fault.
>>
>> With this change, the VA -> IPA translation will only happen for
>> permission faults that are not related to a translation table of a
>> first stage translation.
>>
>> Signed-off-by: Julien Grall <julien.grall@arm.com>
>>
>> ---
>>      Changes in v2:
>>          - Use fsc in the switch in do_trap_data_abort_guest
>> ---
>>   xen/arch/arm/traps.c | 24 ++++++++++++++++++++----
>>   1 file changed, 20 insertions(+), 4 deletions(-)
>>
>> diff --git a/xen/arch/arm/traps.c b/xen/arch/arm/traps.c
>> index ea105f2..83a30fa 100644
>> --- a/xen/arch/arm/traps.c
>> +++ b/xen/arch/arm/traps.c
>> @@ -2382,13 +2382,28 @@ static inline paddr_t get_faulting_ipa(vaddr_t
>> gva)
>>       return ipa;
>>   }
>>
>> +static inline bool hpfar_is_valid(bool s1ptw, uint8_t fsc)
>> +{
>> +    /*
>> +     * HPFAR is valid if one of the following cases are true:
>> +     *  1. the stage 2 fault happen during a stage 1 page table walk
>> +     *  (the bit ESR_EL2.S1PTW is set)
>> +     *  2. the fault was due to a translation fault
>> +     *
>> +     * Note that technically HPFAR is valid for other cases, but they
>> +     * are currently not supported by Xen.
>> +     */
>> +    return s1ptw || (fsc == FSC_FLT_TRANS);
>
> Yes, XEN is not supporting the stage 2 access flag but we should handle
> a stage 2 address size fault.

The function hpfar_is_valid indicates whether the register HPFAR is 
valid. If the function returns false, Xen will use the hardware do the 
translation.

It will only lead to a waste of cycle but this is fine as the address 
size fault is not a hot path for now.

> I think we should do some thing like to below to match ARM ARM.
>
> return s1ptw || (fsc != FSC_FLT_PERM);

This does not match the ARM ARM, with this check you consider that HPFAR 
will be valid for all the fault but permission ones which is not true.

I purposefully choose a white list because it is safer to use the 
hardware to do the translation more often than the invert.

So I don't see why we should handle stage 2 access flag with the current 
Xen. If you still disagree, please explain why with a concrete example.

Regards,
Julien Grall Aug. 18, 2016, 12:02 p.m. UTC | #3
On 17/08/16 21:08, Shanker Donthineni wrote:
>
>
> On 08/17/2016 06:11 AM, Julien Grall wrote:
>> On 17/08/16 03:19, Shanker Donthineni wrote:
>>> Hi Julien,
>>
>> Hello Shanker,
>>
>>> On 07/27/2016 12:09 PM, Julien Grall wrote:
>>>> Translating a VA to a IPA is expensive. Currently, Xen is assuming that
>>>> HPFAR_EL2 is only valid when the stage-2 data/instruction abort
>>>> happened
>>>> during a translation table walk of a first stage translation (i.e S1PTW
>>>> is set).
>>>>
>>>> However, based on the ARM ARM (D7.2.34 in DDI 0487A.j), the register is
>>>> also valid when the data/instruction abort occured for a translation
>>>> fault.
>>>>
>>>> With this change, the VA -> IPA translation will only happen for
>>>> permission faults that are not related to a translation table of a
>>>> first stage translation.
>>>>
>>>> Signed-off-by: Julien Grall <julien.grall@arm.com>
>>>>
>>>> ---
>>>>      Changes in v2:
>>>>          - Use fsc in the switch in do_trap_data_abort_guest
>>>> ---
>>>>   xen/arch/arm/traps.c | 24 ++++++++++++++++++++----
>>>>   1 file changed, 20 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/xen/arch/arm/traps.c b/xen/arch/arm/traps.c
>>>> index ea105f2..83a30fa 100644
>>>> --- a/xen/arch/arm/traps.c
>>>> +++ b/xen/arch/arm/traps.c
>>>> @@ -2382,13 +2382,28 @@ static inline paddr_t get_faulting_ipa(vaddr_t
>>>> gva)
>>>>       return ipa;
>>>>   }
>>>>
>>>> +static inline bool hpfar_is_valid(bool s1ptw, uint8_t fsc)
>>>> +{
>>>> +    /*
>>>> +     * HPFAR is valid if one of the following cases are true:
>>>> +     *  1. the stage 2 fault happen during a stage 1 page table walk
>>>> +     *  (the bit ESR_EL2.S1PTW is set)
>>>> +     *  2. the fault was due to a translation fault
>>>> +     *
>>>> +     * Note that technically HPFAR is valid for other cases, but they
>>>> +     * are currently not supported by Xen.
>>>> +     */
>>>> +    return s1ptw || (fsc == FSC_FLT_TRANS);
>>>
>>> Yes, XEN is not supporting the stage 2 access flag but we should handle
>>> a stage 2 address size fault.
>>
>> The function hpfar_is_valid indicates whether the register HPFAR is
>> valid. If the function returns false, Xen will use the hardware do the
>> translation.
>>
>> It will only lead to a waste of cycle but this is fine as the address
>> size fault is not a hot path for now.
>>
>>> I think we should do some thing like to below to match ARM ARM.
>>>
>>> return s1ptw || (fsc != FSC_FLT_PERM);
>>
>> This does not match the ARM ARM, with this check you consider that
>> HPFAR will be valid for all the fault but permission ones which is not
>> true.
>>
>> I purposefully choose a white list because it is safer to use the
>> hardware to do the translation more often than the invert.
>>
>> So I don't see why we should handle stage 2 access flag with the
>> current Xen. If you still disagree, please explain why with a concrete
>> example.
>>
>
> Agree with you, I have suggested the above change because I saw the same
> check in Linux KVM.
> As per ARM ARM, it should be 'return s1ptw || (fsc == FSC_FLT_TRANS) ||
> (fsc == FSC_FLT_ACCESS) || (fsc == 0x00)';

Feel free to send a patch for this explaining why we would need to have 
the full check. But I don't think it is worth to test every case and 
therefore increasing the number of cycles of this function for faults we 
don't care so far.

Regards,
diff mbox

Patch

diff --git a/xen/arch/arm/traps.c b/xen/arch/arm/traps.c
index ea105f2..83a30fa 100644
--- a/xen/arch/arm/traps.c
+++ b/xen/arch/arm/traps.c
@@ -2382,13 +2382,28 @@  static inline paddr_t get_faulting_ipa(vaddr_t gva)
     return ipa;
 }
 
+static inline bool hpfar_is_valid(bool s1ptw, uint8_t fsc)
+{
+    /*
+     * HPFAR is valid if one of the following cases are true:
+     *  1. the stage 2 fault happen during a stage 1 page table walk
+     *  (the bit ESR_EL2.S1PTW is set)
+     *  2. the fault was due to a translation fault
+     *
+     * Note that technically HPFAR is valid for other cases, but they
+     * are currently not supported by Xen.
+     */
+    return s1ptw || (fsc == FSC_FLT_TRANS);
+}
+
 static void do_trap_instr_abort_guest(struct cpu_user_regs *regs,
                                       const union hsr hsr)
 {
     int rc;
     register_t gva = READ_SYSREG(FAR_EL2);
+    uint8_t fsc = hsr.iabt.ifsc & ~FSC_LL_MASK;
 
-    switch ( hsr.iabt.ifsc & ~FSC_LL_MASK )
+    switch ( fsc )
     {
     case FSC_FLT_PERM:
     {
@@ -2399,7 +2414,7 @@  static void do_trap_instr_abort_guest(struct cpu_user_regs *regs,
             .kind = hsr.iabt.s1ptw ? npfec_kind_in_gpt : npfec_kind_with_gla
         };
 
-        if ( hsr.iabt.s1ptw )
+        if ( hpfar_is_valid(hsr.iabt.s1ptw, fsc) )
             gpa = get_faulting_ipa(gva);
         else
         {
@@ -2434,6 +2449,7 @@  static void do_trap_data_abort_guest(struct cpu_user_regs *regs,
     const struct hsr_dabt dabt = hsr.dabt;
     int rc;
     mmio_info_t info;
+    uint8_t fsc = hsr.dabt.dfsc & ~FSC_LL_MASK;
 
     info.dabt = dabt;
 #ifdef CONFIG_ARM_32
@@ -2442,7 +2458,7 @@  static void do_trap_data_abort_guest(struct cpu_user_regs *regs,
     info.gva = READ_SYSREG64(FAR_EL2);
 #endif
 
-    if ( dabt.s1ptw )
+    if ( hpfar_is_valid(hsr.iabt.s1ptw, fsc) )
         info.gpa = get_faulting_ipa(info.gva);
     else
     {
@@ -2451,7 +2467,7 @@  static void do_trap_data_abort_guest(struct cpu_user_regs *regs,
             return; /* Try again */
     }
 
-    switch ( dabt.dfsc & ~FSC_LL_MASK )
+    switch ( fsc )
     {
     case FSC_FLT_PERM:
     {