Re: [PATCH V3 01/16] machine: anon-alloc option

[David Hildenbrand <david@redhat.com>]

On 04.11.24 18:38, Steven Sistare wrote:
> On 11/4/2024 5:39 AM, David Hildenbrand wrote:
>> On 01.11.24 14:47, Steve Sistare wrote:
>>> Allocate anonymous memory using mmap MAP_ANON or memfd_create depending
>>> on the value of the anon-alloc machine property.  This option applies to
>>> memory allocated as a side effect of creating various devices. It does
>>> not apply to memory-backend-objects, whether explicitly specified on
>>> the command line, or implicitly created by the -m command line option.
>>>
>>> The memfd option is intended to support new migration modes, in which the
>>> memory region can be transferred in place to a new QEMU process, by sending
>>> the memfd file descriptor to the process.  Memory contents are preserved,
>>> and if the mode also transfers device descriptors, then pages that are
>>> locked in memory for DMA remain locked.  This behavior is a pre-requisite
>>> for supporting vfio, vdpa, and iommufd devices with the new modes.
>>
>> A more portable, non-Linux specific variant of this will be using shm,
>> similar to backends/hostmem-shm.c.
>>
>> Likely we should be using that instead of memfd, or try hiding the
>> details. See below.
> 
> For this series I would prefer to use memfd and hide the details.  It's a
> concise (and well tested) solution albeit linux only.  The code you supply
> for posix shm would be a good follow on patch to support other unices.

Unless there is reason to use memfd we should start with the more 
generic POSIX variant that is available even on systems without memfd. 
Factoring stuff out as I drafted does look quite compelling.

I can help with the rework, and send it out separately, so you can focus 
on the "machine toggle" as part of this series.

Of course, if we find out we need the memfd internally instead under 
Linux for whatever reason later, we can use that instead.

But IIUC, the main selling point for memfd are additional features 
(hugetlb, memory sealing) that you aren't even using.

> 
> We could drop
>     -machine anon-alloc=mmap|memfd

Right, the memfd here might be an unnecessary detail. Especially, 
because all things here are mmap'ed ... so I don't quite like this 
interface :)


> and define
>     -machine anon-shared
> 
> as you suggest at the end.

Likely we should remove the "anon" part from the interface as well ... 
hmm ...

We want to instruct QEMU: "all guest RAM that is not explicitly 
specified should be sharable with another process".

"internal-ram-share=true"

"default-ram-share=true"

Maybe we can come up with something even better. But getting rid of the 
"anon" would be great. I think I prefer the latter (below).

> 
>> [...]
>>
>>> @@ -69,6 +70,8 @@
>>>    #include "qemu/pmem.h"
>>> +#include "qapi/qapi-types-migration.h"
>>> +#include "migration/options.h"
>>>    #include "migration/vmstate.h"
>>>    #include "qemu/range.h"
>>> @@ -1849,6 +1852,35 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)
>>>                    qemu_mutex_unlock_ramlist();
>>>                    return;
>>>                }
>>> +
>>> +        } else if (current_machine->anon_alloc == ANON_ALLOC_OPTION_MEMFD &&
>>> +                   !object_dynamic_cast(new_block->mr->parent_obj.parent,
>>> +                                        TYPE_MEMORY_BACKEND)) {
>>
>> This looks a bit and hackish,
> 
> OK. I can revert parts of the previous version which passed in RAM_SHARED from
> various call sites to request anonymous shared memory:
>     https://lore.kernel.org/qemu-devel/1714406135-451286-18-git-send-email-steven.sistare@oracle.com
> See the various sites that do
>       uint32_t flags = current_machine->memfd_alloc ? RAM_SHARED : 0;
> Does that look OK to you?

That's one option, or we just handle it in qemu_ram_alloc_internal() as 
I drafted below.

Or we simply have another interface to allocate this "default RAM that 
does not come from a memory backend and is subject to the global 
toggle", and hide that detail (conditionally setting RAM_SHARED) in there.

> 
>> and I don't think ram_block_add() is the right
>> place where this should be. It should likely happen in the caller.
> 
> I agree, but I received no feedback when I proposed to refactor allocation
> vs ram_block_add, so I dropped them to simplify the live update review.
> These refactor but do not change functionality.  Are you OK with something
> like this?  Is this overkill?
> 

Probably overkill :)

> https://lore.kernel.org/qemu-devel/1714406135-451286-1-git-send-email-steven.sistare@oracle.com/
>     physmem: ram_block_create
>     physmem: hoist guest_memfd creation
>     physmem: hoist host memory allocation
> 
>> We already do have two ways of allocating "shared anonymous memory":
>>
>> (1) memory-backend-ram,share=on
>> (2) memory-backend-shm
>>
>> (2) gives us an fd as it uses shm_open(), (1) doesn't give us an fd as it
>> uses MAP_ANON|MAP_SHARED. (1) is really only a corner case use case [1].
>>
>> [there is also Linux specific memfd, which gives us more flexibility with
>> hugetlb etc, but for the purpose here shm should likely be sufficient?]
>>
>> So why not make (1) behave like (2) and move that handling into
>> qemu_ram_alloc_internal(), from where we can easily enable it using a
>> new RMA_SHARED flag? So as a first step, something like:
> 
> I prefer that, and an earlier version did so, but only if anon-alloc==memfd.
> 
> To be clear, do you propose that memory-backend-ram,shared=on unconditionally
> mmap fd-based shared memory, independently of the setting of anon-alloc?
> And drop the MAP_ANON|MAP_SHARED possibility?

Yes, as done in my draft patch. MAP_ANON|MAP_SHARED was primarily a hack 
to make this RDMA thingy fly that could not deal with anonymous memory, 
and we didn't have

memory-backend-ram,share=on was added via 
06329ccecfa022494fdba288b3ab5bcb8dff4159 before
memory-backend-memfd was added via dbb9e0f40d7d561dcffcf7e41ac9f6a5ec90e5b5

Both ended up in the same QEMU release.

So likely memory-backend-ram,share=on could just have used 
memory-backend-memfd if it would have been available earlier, at least 
on Linux ...


But, it looks like the use case for memory-backend-ram,share=on does no 
longer even exist, because

commit 1dfd42c4264bbf47415a9e73f0d6b4e6a7cd7393
Author: Philippe Mathieu-Daudé <philmd@linaro.org>
Date:   Thu Mar 28 12:53:00 2024 +0100

     hw/rdma: Remove deprecated pvrdma device and rdmacm-mux helper

Removed that mremap() from the code base.

So we can change how memory-backend-ram,share=on is implemented 
internally, as long as it keeps on working in a similar way.

If "memory-backend-ram,share=on" will be the same as specifying 
"default-ram-share=on", that would actually be quite nice ... no need to 
bring in memfds at all as long as we only want some memory with an fd to 
share with other processes.

> 
> Or, do you propose that for memory-backend-ram,shared=on:
>     if anon-shared
>       mmap fd
>     else
>        MAP_ANON|MAP_SHARED


My suggestion would be to unconditionally use shm (which is available 
even on kernels without memfd support; if required use memfd first and 
fallback to shmem) as in the patch I drafted.

> 
> The former is simpler from a user documentation point of view, but either
> works for me.  I could stop listing memory-backend-ram  as an exception in
> the docs, which currently state:
>     #     Memory-backend objects must have the share=on attribute, but
>     #     memory-backend-epc and memory-backend-ram are not supported.

Likely that was never updated to document the memory-backend-ram use case.

> 
> [...]
>>
>> Then, you only need a machine option to say "anon-shared", to make all
>> anonymous memory sharable between processes. All it would do is setting
>> the RAM_SHARED flag in qemu_ram_alloc_internal() when reasonable
>> (!(ram_flags & RAM_PREALLOC)).
>>
>> To handle "memory-backend-ram,share=off", can we find a way to bail out if
>> memory-backend-ram,share=off was used while the machine option "anon-shared"
>> would be active?
> 
> In later patches I install migration blockers for various conditions, including
> when a ram block does not support CPR.

Good!

> 
>> Or just document that the "anon-shared" will win?
> 
> IMO a blocker is sufficient.
> 
> I think you are also suggesting that an unadorned "memory-backend-ram"
> specification (with implicit shared=off), plus anon-shared, should cause
> shared anon to be allocated:
>     "you only need a machine option to say "anon-shared", to make all anonymous
>      memory sharable"
> 
> I did that previously, and Peter objected, saying the explicit anon-shared
> should not override the implicit shared=off.

Yes, it's better if we can detect that somehow. There should be easy 
ways to make that work, so I wouldn't worry about that.

-- 
Cheers,

David / dhildenb