On Sat, Apr 22, 2023 at 8:46 AM Akihiko Odaki <akihiko.od...@gmail.com> wrote: > > 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?
It looks like dma_memory_map(..) gives a host virtual address given an guest DMA address? If so, that won't work. We need to give the guest a pointer to the host's GPU memory. I took a look at what the Android Emulator has been doing: it largely bypasses the MemoryRegion API and defines generic map/unmap functions: https://android.googlesource.com/platform/external/qemu/+/emu-master-dev/hw/display/virtio-gpu-3d.c#473 https://android.googlesource.com/platform/external/qemu/+/emu-master-dev/exec.c#2472 https://android.googlesource.com/platform/external/qemu/+/emu-master-dev/accel/kvm/kvm-all.c#1787 This is kind of what crosvm does too. That seems the simplest way to do what we need. Any objections to this approach or alternative methods? > > > > > [a] https://qemu.readthedocs.io/en/latest/devel/memory.html > > > > Signed-off-by: Gurchetan Singh <gurchetansi...@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,
