From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from draig.lan ([185.126.160.19]) by smtp.gmail.com with ESMTPSA id a640c23a62f3a-ade1dc7b202sm91694966b.161.2025.06.06.03.16.36 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Fri, 06 Jun 2025 03:16:37 -0700 (PDT) Received: from draig (localhost [IPv6:::1]) by draig.lan (Postfix) with ESMTP id 6905C5F7E1; Fri, 06 Jun 2025 11:16:36 +0100 (BST) From: =?utf-8?Q?Alex_Benn=C3=A9e?= To: Akihiko Odaki Cc: qemu-devel@nongnu.org, Sriram Yagnaraman , "Michael S. Tsirkin" , Dmitry Osipenko , Paolo Bonzini , Peter Maydell , John Snow , =?utf-8?Q?Marc-Andr=C3=A9?= Lureau , Pierrick Bouvier , Peter Xu , Fabiano Rosas , qemu-arm@nongnu.org, Thomas Huth , Alexandre Iooss , Gustavo Romero , Markus Armbruster , David Hildenbrand , Daniel P. =?utf-8?Q?Berrang=C3=A9?= , Laurent Vivier , Philippe =?utf-8?Q?Mathieu-Daud=C3=A9?= , Mahmoud Mandour , Manos Pitsidianakis , qemu-stable@nongnu.org Subject: Re: [PATCH v4 09/17] hw/display: re-arrange memory region tracking In-Reply-To: <5f91c8a2-06ce-45f8-97bd-0602a52e0d21@daynix.com> (Akihiko Odaki's message of "Fri, 6 Jun 2025 18:40:06 +0900") References: <20250603110204.838117-1-alex.bennee@linaro.org> <20250603110204.838117-10-alex.bennee@linaro.org> <1a86b86d-145a-44fc-9f87-2804767fb109@daynix.com> <87o6v2764e.fsf@draig.linaro.org> <5f91c8a2-06ce-45f8-97bd-0602a52e0d21@daynix.com> User-Agent: mu4e 1.12.11; emacs 30.1 Date: Fri, 06 Jun 2025 11:16:36 +0100 Message-ID: <87tt4t41kr.fsf@draig.linaro.org> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable X-TUID: cVqSWJyz04QB Akihiko Odaki writes: > On 2025/06/05 20:57, Alex Benn=C3=A9e wrote: >> Akihiko Odaki writes: >>=20 >>> On 2025/06/03 20:01, Alex Benn=C3=A9e wrote: >>>> QOM objects can be embedded in other QOM objects and managed as part >>>> of their lifetime but this isn't the case for >>>> virtio_gpu_virgl_hostmem_region. However before we can split it out we >>>> need some other way of associating the wider data structure with the >>>> memory region. >>>> Fortunately MemoryRegion has an opaque pointer. This is passed down >>>> to >>>> MemoryRegionOps for device type regions but is unused in the >>>> memory_region_init_ram_ptr() case. Use the opaque to carry the >>>> reference and allow the final MemoryRegion object to be reaped when >>>> its reference count is cleared. >>>> Signed-off-by: Manos Pitsidianakis >>>> Reviewed-by: Philippe Mathieu-Daud=C3=A9 >>>> Signed-off-by: Alex Benn=C3=A9e >>>> Message-Id: <20250410122643.1747913-2-manos.pitsidianakis@linaro.org> >>>> Cc: qemu-stable@nongnu.org >>>> --- >>>> include/system/memory.h | 1 + >>>> hw/display/virtio-gpu-virgl.c | 23 ++++++++--------------- >>>> 2 files changed, 9 insertions(+), 15 deletions(-) >>>> diff --git a/include/system/memory.h b/include/system/memory.h >>>> index fc35a0dcad..90715ff44a 100644 >>>> --- a/include/system/memory.h >>>> +++ b/include/system/memory.h >>>> @@ -784,6 +784,7 @@ struct MemoryRegion { >>>> DeviceState *dev; >>>> const MemoryRegionOps *ops; >>>> + /* opaque data, used by backends like @ops */ >>>> void *opaque; >>>> MemoryRegion *container; >>>> int mapped_via_alias; /* Mapped via an alias, container might b= e NULL */ >>>> diff --git a/hw/display/virtio-gpu-virgl.c b/hw/display/virtio-gpu-vir= gl.c >>>> index 145a0b3879..71a7500de9 100644 >>>> --- a/hw/display/virtio-gpu-virgl.c >>>> +++ b/hw/display/virtio-gpu-virgl.c >>>> @@ -52,17 +52,11 @@ virgl_get_egl_display(G_GNUC_UNUSED void *cookie) >>>> #if VIRGL_VERSION_MAJOR >=3D 1 >>>> struct virtio_gpu_virgl_hostmem_region { >>>> - MemoryRegion mr; >>>> + MemoryRegion *mr; >>>> struct VirtIOGPU *g; >>>> bool finish_unmapping; >>>> }; >>>> -static struct virtio_gpu_virgl_hostmem_region * >>>> -to_hostmem_region(MemoryRegion *mr) >>>> -{ >>>> - return container_of(mr, struct virtio_gpu_virgl_hostmem_region, m= r); >>>> -} >>>> - >>>> static void virtio_gpu_virgl_resume_cmdq_bh(void *opaque) >>>> { >>>> VirtIOGPU *g =3D opaque; >>>> @@ -73,14 +67,12 @@ static void virtio_gpu_virgl_resume_cmdq_bh(void *= opaque) >>>> static void virtio_gpu_virgl_hostmem_region_free(void *obj) >>>> { >>>> MemoryRegion *mr =3D MEMORY_REGION(obj); >>>> - struct virtio_gpu_virgl_hostmem_region *vmr; >>>> + struct virtio_gpu_virgl_hostmem_region *vmr =3D mr->opaque; >>>> VirtIOGPUBase *b; >>>> VirtIOGPUGL *gl; >>>> - vmr =3D to_hostmem_region(mr); >>>> - vmr->finish_unmapping =3D true; >>>> - >>>> b =3D VIRTIO_GPU_BASE(vmr->g); >>>> + vmr->finish_unmapping =3D true; >>>> b->renderer_blocked--; >>>> /* >>>> @@ -118,8 +110,8 @@ virtio_gpu_virgl_map_resource_blob(VirtIOGPU *g, >>>> vmr =3D g_new0(struct virtio_gpu_virgl_hostmem_region, 1); >>>> vmr->g =3D g; >>>> + mr =3D g_new0(MemoryRegion, 1); >>> >>> This patch does nothing more than adding a separate allocation for >>> MemoryRegion. Besides there is no corresponding g_free(). This patch >>> can be simply dropped. >> As the patch says the MemoryRegion is now free'd when it is >> de-referenced. Do you have a test case showing it leaking? > > "De-referenced" is confusing and sounds like pointer dereferencing. > > OBJECT(mr)->free, which has virtio_gpu_virgl_hostmem_region_free() as > its value, will be called to free mr when the references of mr are > removed. This patch however does not add a corresponding g_free() call > to virtio_gpu_virgl_hostmem_region_free(), leaking mr. > > AddressSanitizer will catch the memory leak. Example invocation? I ran the AddressSantizier against all the virtio-gpu tests yesterday and it did not complain. > > Regards, > Akihiko Odaki --=20 Alex Benn=C3=A9e Virtualisation Tech Lead @ Linaro