Re: [RFC PATCH 12/13] HACK: use memory region API to inject memory to guest

[Akihiko Odaki <akihiko.odaki@gmail.com>]

On 2023/04/21 10:12, Gurchetan Singh wrote:
> I just copied the patches that have been floating around that do
> this, but it doesn't seem to robustly work.  This current
> implementation is probably good enough to run vkcube or simple
> apps, but whenever a test starts to aggressively map/unmap memory,
> things do explode on the QEMU side.
> 
> A simple way to reproduce is run:
> 
> ./deqp-vk --deqp-case=deqp-vk --deqp-case=dEQP-VK.memory.mapping.suballocation.*
> 
> You should get stack traces that sometimes look like this:
> 
> 0  __pthread_kill_implementation (no_tid=0, signo=6, threadid=140737316304448) at ./nptl/pthread_kill.c:44
> 1  __pthread_kill_internal (signo=6, threadid=140737316304448) at ./nptl/pthread_kill.c:78
> 2  __GI___pthread_kill (threadid=140737316304448, signo=signo@entry=6) at ./nptl/pthread_kill.c:89
> 3  0x00007ffff7042476 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
> 4  0x00007ffff70287f3 in __GI_abort () at ./stdlib/abort.c:79
> 5  0x00007ffff70896f6 in __libc_message (action=action@entry=do_abort, fmt=fmt@entry=0x7ffff71dbb8c "%s\n") at ../sysdeps/posix/libc_fatal.c:155
> 6  0x00007ffff70a0d7c in malloc_printerr (str=str@entry=0x7ffff71de7b0 "double free or corruption (out)") at ./malloc/malloc.c:5664
> 7  0x00007ffff70a2ef0 in _int_free (av=0x7ffff7219c80 <main_arena>, p=0x555557793e00, have_lock=<optimized out>) at ./malloc/malloc.c:4588
> 8  0x00007ffff70a54d3 in __GI___libc_free (mem=<optimized out>) at ./malloc/malloc.c:3391
> 9  0x0000555555d65e7e in phys_section_destroy (mr=0x555557793e10) at ../softmmu/physmem.c:1003
> 10 0x0000555555d65ed0 in phys_sections_free (map=0x555556d4b410) at ../softmmu/physmem.c:1011
> 11 0x0000555555d69578 in address_space_dispatch_free (d=0x555556d4b400) at ../softmmu/physmem.c:2430
> 12 0x0000555555d58412 in flatview_destroy (view=0x5555572bb090) at ../softmmu/memory.c:292
> 13 0x000055555600fd23 in call_rcu_thread (opaque=0x0) at ../util/rcu.c:284
> 14 0x00005555560026d4 in qemu_thread_start (args=0x5555569cafa0) at ../util/qemu-thread-posix.c:541
> 15 0x00007ffff7094b43 in start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:442
> 16 0x00007ffff7126a00 in clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81
> 
> or this:
> 
> 0x0000555555e1dc80 in object_unref (objptr=0x6d656d3c6b6e696c) at ../qom/object.c:1198
> 1198        g_assert(obj->ref > 0);
> (gdb) bt
> 0  0x0000555555e1dc80 in object_unref (objptr=0x6d656d3c6b6e696c) at ../qom/object.c:1198
> 1  0x0000555555d5cca5 in memory_region_unref (mr=0x5555572b9e20) at ../softmmu/memory.c:1799
> 2  0x0000555555d65e47 in phys_section_destroy (mr=0x5555572b9e20) at ../softmmu/physmem.c:998
> 3  0x0000555555d65ec7 in phys_sections_free (map=0x5555588365c0) at ../softmmu/physmem.c:1011
> 4  0x0000555555d6956f in address_space_dispatch_free (d=0x5555588365b0) at ../softmmu/physmem.c:2430
> 5  0x0000555555d58409 in flatview_destroy (view=0x555558836570) at ../softmmu/memory.c:292
> 6  0x000055555600fd1a in call_rcu_thread (opaque=0x0) at ../util/rcu.c:284
> 7  0x00005555560026cb in qemu_thread_start (args=0x5555569cafa0) at ../util/qemu-thread-posix.c:541
> 8  0x00007ffff7094b43 in start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:442
> 9  0x00007ffff7126a00 in clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81
> 
> The reason seems to be memory regions are handled on a different
> thread than the virtio-gpu thread, and that inevitably leads to
> raciness.  The memory region docs[a] generally seems to dissuade this:
> 
> "In order to do this, as a general rule do not create or destroy
>   memory regions dynamically during a device’s lifetime, and only
>   call object_unparent() in the memory region owner’s instance_finalize
>   callback. The dynamically allocated data structure that contains
>   the memory region then should obviously be freed in the
>   instance_finalize callback as well."
> 
> Though instance_finalize is called before device destruction, so
> storing the memory until then is unlikely to be an option.  The
> tests do pass when virtio-gpu doesn't free the memory, but
> progressively the guest becomes slower and then OOMs.
> 
> Though the api does make an exception:
> 
> "There is an exception to the above rule: it is okay to call
> object_unparent at any time for an alias or a container region. It is
> therefore also okay to create or destroy alias and container regions
> dynamically during a device’s lifetime."
> 
> I believe we are trying to create a container subregion, but that's
> still failing?  Are we doing it right?  Any memory region experts
> here can help out?  The other revelant patch in this series
> is "virtio-gpu: hostmem".

Perhaps dma_memory_map() is what you want?

> 
> [a] https://qemu.readthedocs.io/en/latest/devel/memory.html
> 
> Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
> ---
>   hw/display/virtio-gpu-rutabaga.c | 14 ++++++++++++++
>   1 file changed, 14 insertions(+)
> 
> diff --git a/hw/display/virtio-gpu-rutabaga.c b/hw/display/virtio-gpu-rutabaga.c
> index 5fd1154198..196267aac2 100644
> --- a/hw/display/virtio-gpu-rutabaga.c
> +++ b/hw/display/virtio-gpu-rutabaga.c
> @@ -159,6 +159,12 @@ static int32_t rutabaga_handle_unmap(VirtIOGPU *g,
>       GET_VIRTIO_GPU_GL(g);
>       GET_RUTABAGA(virtio_gpu);
>   
> +    memory_region_transaction_begin();
> +    memory_region_set_enabled(&res->region, false);
> +    memory_region_del_subregion(&g->parent_obj.hostmem, &res->region);
> +    object_unparent(OBJECT(&res->region));
> +    memory_region_transaction_commit();
> +
>       res->mapped = NULL;
>       return rutabaga_resource_unmap(rutabaga, res->resource_id);
>   }
> @@ -671,6 +677,14 @@ rutabaga_cmd_resource_map_blob(VirtIOGPU *g,
>       result = rutabaga_resource_map(rutabaga, mblob.resource_id, &mapping);
>       CHECK_RESULT(result, cmd);
>   
> +    memory_region_transaction_begin();
> +    memory_region_init_ram_device_ptr(&res->region, OBJECT(g), NULL,
> +                                      mapping.size, (void *)mapping.ptr);
> +    memory_region_add_subregion(&g->parent_obj.hostmem, mblob.offset,
> +                                &res->region);
> +    memory_region_set_enabled(&res->region, true);
> +    memory_region_transaction_commit();
> +
>       memset(&resp, 0, sizeof(resp));
>       resp.hdr.type = VIRTIO_GPU_RESP_OK_MAP_INFO;
>       result = rutabaga_resource_map_info(rutabaga, mblob.resource_id,