[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH 3/3] x86/mem_sharing: make fork_reset more configurable
- To: Jan Beulich <jbeulich@xxxxxxxx>
- From: Tamas K Lengyel <tamas@xxxxxxxxxxxxx>
- Date: Fri, 25 Mar 2022 06:48:42 -0400
- Cc: Tamas K Lengyel <tamas.lengyel@xxxxxxxxx>, Xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxxx>, Wei Liu <wl@xxxxxxx>, Anthony PERARD <anthony.perard@xxxxxxxxxx>, Juergen Gross <JGross@xxxxxxxx>, Andrew Cooper <andrew.cooper3@xxxxxxxxxx>, George Dunlap <george.dunlap@xxxxxxxxxx>, Julien Grall <julien@xxxxxxx>, Stefano Stabellini <sstabellini@xxxxxxxxxx>, Alexandru Isaila <aisaila@xxxxxxxxxxxxxxx>, Petre Pircalabu <ppircalabu@xxxxxxxxxxxxxxx>, Roger Pau Monné <roger.pau@xxxxxxxxxx>
- Delivery-date: Fri, 25 Mar 2022 10:49:07 +0000
- List-id: Xen developer discussion <xen-devel.lists.xenproject.org>
On 24.03.2022 18:02, Tamas K Lengyel wrote:
> On Thu, Mar 24, 2022 at 12:44 PM Roger Pau Monné <roger.pau@xxxxxxxxxx> wrote:
>>
>> On Thu, Mar 24, 2022 at 12:22:49PM -0400, Tamas K Lengyel wrote:
>>> On Thu, Mar 24, 2022 at 12:04 PM Roger Pau Monné <roger.pau@xxxxxxxxxx> wrote:
>>>>
>>>> On Thu, Mar 24, 2022 at 11:52:38AM -0400, Tamas K Lengyel wrote:
>>>>> On Thu, Mar 24, 2022 at 11:46 AM Roger Pau Monné <roger.pau@xxxxxxxxxx> wrote:
>>>>>>
>>>>>> On Tue, Mar 22, 2022 at 01:41:39PM -0400, Tamas K Lengyel wrote:
>>>>>>> diff --git a/xen/include/public/memory.h b/xen/include/public/memory.h
>>>>>>> index 208d8dcbd9..30ce23c5a7 100644
>>>>>>> --- a/xen/include/public/memory.h
>>>>>>> +++ b/xen/include/public/memory.h
>>>>>>> @@ -541,12 +541,14 @@ struct xen_mem_sharing_op {
>>>>>>> uint32_t gref; /* IN: gref to debug */
>>>>>>> } u;
>>>>>>> } debug;
>>>>>>> - struct mem_sharing_op_fork { /* OP_FORK */
>>>>>>> + struct mem_sharing_op_fork { /* OP_FORK/_RESET */
>>>>>>> domid_t parent_domain; /* IN: parent's domain id */
>>>>>>> /* These flags only makes sense for short-lived forks */
>>>>>>> #define XENMEM_FORK_WITH_IOMMU_ALLOWED (1u << 0)
>>>>>>> #define XENMEM_FORK_BLOCK_INTERRUPTS (1u << 1)
>>>>>>> #define XENMEM_FORK_SKIP_SPECIAL_PAGES (1u << 2)
>>>>>>> +#define XENMEM_FORK_RESET_STATE (1u << 3)
>>>>>>> +#define XENMEM_FORK_RESET_MEMORY (1u << 4)
>>>>>>> uint16_t flags; /* IN: optional settings */
>>>>>>> uint32_t pad; /* Must be set to 0 */
>>>>>>> } fork;
>>>>>>> diff --git a/xen/include/public/vm_event.h b/xen/include/public/vm_event.h
>>>>>>> index bb003d21d0..81c2ee28cc 100644
>>>>>>> --- a/xen/include/public/vm_event.h
>>>>>>> +++ b/xen/include/public/vm_event.h
>>>>>>> @@ -127,6 +127,14 @@
>>>>>>> * Reset the vmtrace buffer (if vmtrace is enabled)
>>>>>>> */
>>>>>>> #define VM_EVENT_FLAG_RESET_VMTRACE (1 << 13)
>>>>>>> +/*
>>>>>>> + * Reset the VM state (if VM is fork)
>>>>>>> + */
>>>>>>> +#define VM_EVENT_FLAG_RESET_FORK_STATE (1 << 14)
>>>>>>> +/*
>>>>>>> + * Remove unshared entried from physmap (if VM is fork)
>>>>>>> + */
>>>>>>> +#define VM_EVENT_FLAG_RESET_FORK_MEMORY (1 << 15)
>>>>>>
>>>>>> I'm confused about why two different interfaces are added to do this
>>>>>> kind of selective resets, one to vm_event and one to xenmem_fork?
>>>>>>
>>>>>> I thin k the natural place for the option to live would be
>>>>>> XENMEM_FORK?
>>>>>
>>>>> Yes, that's the natural place for it. But we are adding it to both for
>>>>> a reason. In our use-case the reset operation will happen after a
>>>>> vm_event is received to which we already must send a reply. Setting
>>>>> the flag on the vm_event reply saves us having to issue an extra memop
>>>>> hypercall afterwards.
>>>>
>>>> Can you do a multicall and batch both operations in a single
>>>> hypercall?
>>>>
>>>> That would seem more natural than adding duplicated interfaces.
>>>
>>> Not in a straight forward way, no. There is no exposed API in libxc to
>>> do a multicall. Even if that was an option it is still easier for me
>>> to just flip a bit in the response field than having to construct a
>>> whole standalone hypercall structure to be sent as part of a
>>> multicall.
>>
>> Right, I can see it being easier, but it seems like a bad choice from
>> an interface PoV. You are the maintainer of both subsystems, but it
>> would seem to me it's in your best interest to try to keep the
>> interfaces separated and clean.
>>
>> Would it be possible for the reset XENMEM_FORK op to have the side
>> effect of performing what you would instead do with the vm_event
>> hypercall?
>
> Yes, the event response is really just an event channel signal to Xen,
> so the memop hypercall could similarly encode the "now check the
> vm_event response" as an optional field. But why is that any better
> than the current event channel route processing the vm_response
> encoding the "now do these ops on the fork"?
Well, as Roger said: Less duplication in the interface.
No, you would just duplicate something else instead, ie. the event channel hypercall.
> We already have a bunch of different operations you can encode in the
> vm_event response field, so it reduces the complexity on the toolstack
> side since I don't have to switch around which hypercall I need to
> issue depending on what extra ops I want to put into a single
> hypercall.
The two goals need to be weighed against one another; for the moment
I think I'm with Roger aiming at a clean interface.
It may look like that from the Xen side but from the toolstack side this is actually the cleanest way to achieve what we need. The vm_event interfaces are already strongly integrated with both the mem_sharing and mem_paging subsystems so nothing is gained by now for no reason trying to keep them separate. So I strongly disagree with this suggestion and I'm going to keep it as-is. I appreciate the feedback nevertheless.
Tamas
|