[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Xen-devel] [PATCH v2 2/5] vm_event: Implement ARM SMC events
> On 03/05/2016 19:48, Tamas K Lengyel wrote:
>>
>>
>>
>> On Tue, May 3, 2016 at 5:31 AM, Julien Grall <julien.grall@xxxxxxx
>> <mailto:julien.grall@xxxxxxx>> wrote:
>> On 29/04/16 19:07, Tamas K Lengyel wrote:
>>
>> The ARM SMC instructions are already configured to trap to Xen
>> by default. In
>> this patch we allow a user-space process in a privileged domain
>> to receive
>> notification of when such event happens through the vm_event
>> subsystem by
>> introducing the PRIVILEGED_CALL type.
>>
>> Signed-off-by: Tamas K Lengyel <tamas@xxxxxxxxxxxxx
>> <mailto:tamas@xxxxxxxxxxxxx>>
>>
>> ---
>> Cc: Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx
>> <mailto:rcojocaru@xxxxxxxxxxxxxxx>>
>> Cc: Ian Jackson <ian.jackson@xxxxxxxxxxxxx
>> <mailto:ian.jackson@xxxxxxxxxxxxx>>
>> Cc: Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx
>> <mailto:stefano.stabellini@xxxxxxxxxxxxx>>
>> Cc: Wei Liu <wei.liu2@xxxxxxxxxx <mailto:wei.liu2@xxxxxxxxxx>>
>> Cc: Julien Grall <julien.grall@xxxxxxx
>> <mailto:julien.grall@xxxxxxx>>
>> Cc: Keir Fraser <keir@xxxxxxx <mailto:keir@xxxxxxx>>
>> Cc: Jan Beulich <jbeulich@xxxxxxxx <mailto:jbeulich@xxxxxxxx>>
>> Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx
>> <mailto:andrew.cooper3@xxxxxxxxxx>>
>>
>>
>> v2: introduce VM_EVENT_REASON_PRIVELEGED_CALL
>> aarch64 support
>> ---
>> MAINTAINERS | 6 +-
>> tools/libxc/include/xenctrl.h | 2 +
>> tools/libxc/xc_monitor.c | 26 +++++++-
>> tools/tests/xen-access/xen-access.c | 31 ++++++++-
>> xen/arch/arm/Makefile | 2 +
>> xen/arch/arm/monitor.c | 80 +++++++++++++++++++++++
>> xen/arch/arm/traps.c | 20 +++++-
>> xen/arch/arm/vm_event.c | 127
>> ++++++++++++++++++++++++++++++++++++
>> xen/arch/x86/hvm/event.c | 2 +
>> xen/common/vm_event.c | 1 -
>> xen/include/asm-arm/domain.h | 5 ++
>> xen/include/asm-arm/monitor.h | 20 ++----
>> xen/include/asm-arm/vm_event.h | 16 ++---
>> xen/include/public/domctl.h | 1 +
>> xen/include/public/vm_event.h | 27 ++++++++
>> 15 files changed, 333 insertions(+), 33 deletions(-)
>> create mode 100644 xen/arch/arm/monitor.c
>> create mode 100644 xen/arch/arm/vm_event.c
>>
>>
>> This patch is doing lots of things:
>> - Add support for monitoring
>> - Add support for vm_event
>> - Monitor SMC
>> - Move common code to arch specific code
>>
>> As far as I can tell, all are distinct from each other. Can you
>> please split this patch in smaller ones?
>>
>>
>> While I can split off some parts into separate patches, like
>> getting/setting ARM registers through VM events and the tools patches,
>> the other components are pretty tightly coupled and don't actually make
>> sense to split them. For example, enabling a monitor domctl for an event
>> without the VM event components doesn't make much sense. Vice verse for
>> the vm_event parts without being able to enable them.
>
>
> Well, the commit message does not mention half of the changes of this patch. Some changes like moving common code to arch specific code clearly needs explanation. It is the same for the fact that you only present a reduce set of registers to vm event for AArch64.
This IMHO is not outstanding, it's the same on x86.
>
> In any case, there is too many logical changes in this patch, which makes difficult to review it. So please split this patch in smaller chunk.
Sure, I already split the parts I mentioned in the previous message.
>
> [...]
>
>
>> + if ( current->domain->arch.monitor.privileged_call_enabled )
>> + {
>> + rc = monitor_smc(regs);
>> + }
>>
>>
>> The bracket are not necessary.
>>
>>
>> Ack.
>>
>>
>> +
>> + if ( rc != 1 )
>>
>>
>> I think the code would be clearer if you introduce a define for "1".
>>
>>
>> Maybe not a define but calling the variable "handled" as we do on x86
>> would be more descriptive.
>
>
> IHMO, "handled" infers that the variable is boolean. This is not the case here as you could have negative value.
It may be but thats the convention we have for this on x86 so symmetry is better then introducing a new define just for the ARM case.
>
>
>>
>>
>> + {
>> + GUEST_BUG_ON(!psr_mode_is_32bit(regs->cpsr));
>>
>>
>> This check cannot work for AArch64 guest.
>>
>>
>> For HSR_EC_SMC32 there is already
>> GUEST_BUG_ON(!psr_mode_is_32bit(regs->cpsr)); and for HSR_EC_SMC64 there
>> is GUEST_BUG_ON(psr_mode_is_32bit(regs->cpsr)); before calling
>> do_trap_smc. So are you saying that check is wrong for AArch64 as it is
>> right now in unstable? Also, is there any reason those checks are
>> opposite of each other?
>
>
> No, I am saying that your check is wrong here. psr_mode_is_32bit returns true if the guest was running in 32-bit mode, else false when running in 64-bit mode.
>
> The check are opposites each other because the exception SMC64 can only be taken from an AArch64 state, and SMC32 from an AArch32 State.
Ok, got it.
>
> [...]
>
>
>> AArch64 provides 31 generate-purpose registers. Although, x29 and
>> x30 are respectively used for fp and lr.
>>
>>
>> For vm_event it's not necessary to get all registers, rather it's just a
>> handful of selection that may be especially "useful" for introspection.
>
>
> How did you decide that only the first to xN registers are useful? It would be valid to have an SMC call using x20 for an arguments.
>
> Similarly, the hypercall convention for AArch64 makes use of x16 which is not exposed to the vm event subsystem.
Certainly, as I said, if a future application needs other registers to be sent here and can justify it, this can be adjusted.
>
>
>> It's also important not to fill up the vm_event monitor ring with huge
>> request/response structs so even on x86 we only have a subset of the
>> registers. As right now there are no applications for aarch64, it's only
>> a guess of what registers would be "useful" and may be adjusted in
>> future versions as we start to have applications using this.
>
>
> Guessing the set of useful registers is usually not a good idea (see why above).
>
Remember, the subscriber can always get/set the full set of registers when it needs to, so completeness here is not critical. You are missing the point that the space on the ring is limited and it can fill up fast when the full vCPU context is pushed on it for all events. Right now the x86 side is still larger so we have some room for additional registers to be sent when users of this system have a better view into what they find important. Which is IMHO not now.
Thanks,
Tamas
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel
|
