On 3/8/19 12:22 PM, Philippe Mathieu-Daudé wrote:
> On 3/8/19 12:04 PM, Laszlo Ersek wrote:
>> Hi Phil,
>>
>> On 03/08/19 02:32, Philippe Mathieu-Daudé wrote:
>>> Due to the contract interface of fw_cfg_add_file(), the
>>> 'reboot_timeout' data has to be valid for the lifetime of the
>>> FwCfg object. For this reason it is copied on the heap with
>>> memdup().
>>>
>>> The object state, 'FWCfgState', is also meant to be valid during the
>>> lifetime of the object.
>>> Move the 'reboot_timeout' in FWCfgState to achieve the same purpose.
>>> Doing so we avoid a memory leak.
>>>
>>> Signed-off-by: Philippe Mathieu-Daudé <phi...@redhat.com>
>>> ---
>>> hw/nvram/fw_cfg.c | 4 +++-
>>> include/hw/nvram/fw_cfg.h | 2 ++
>>> 2 files changed, 5 insertions(+), 1 deletion(-)
>>
>> Currently, there is no memory leak. Right now, the leak is theoretical,
>> and it would depend on the fw_cfg object being actually destroyed.
>
> Actually my first motivation came while using valgrind, there are a
> bunch of warnings related to the fw_cfg device.
> This device is not hotpluggable however, and we don't test it in the
> device-introspect-test.
IOW this device makes finding memory leaks in other introspectable
devices harder.
>> I think armoring the fw_cfg implementation for such lifetime actions is
>> valuable. But, that definitely belongs to its own series, in my opinion.
>>
>> In the "hw/nvram/fw_cfg.c" file, I count:
>>
>> (a) two "specific purpose" g_memdup() calls, namely in
>> fw_cfg_bootsplash() and in fw_cfg_reboot();
>>
>> (b) one "generic purpose" g_memdup() call, namely in fw_cfg_add_string();
>>
>> (c) two "generic purpose" g_malloc() calls, namely in fw_cfg_add_i16(),
>> fw_cfg_add_i32(), and fw_cfg_add_i64(). (The one in fw_cfg_modify_i16()
>> does not matter here because the previous blob is freed in that function.)
>>
>> Your series deals with (a), namely with fw_cfg_reboot() in this patch,
>> and with fw_cfg_bootsplash() in the next one.
>>
>> Your series deals with neither (b) nor (c). The
>
> I did a PoC of (b) and (c) but it is a more invasive patchset indeed.
>
>> fw_cfg_add_(string|i16|i32|i64) functions are called from a bunch of
>> places however, so if we really intend *not* to leak those copies upon
>> fw_cfg destruction, then we'll have to track all of them dynamically, in
>> a list for example.
>
> I haven't think of using a list.
>
>> (And that necessitates a separate series for this topic even more.)
>
> OK.
>
>> In turn, once we add dynamic tracking, for those blobs that the
>> fw_cfg_add_(string|i16|i32|i64) functions allocate internally -- as they
>> are advertized to do --, then we might as well use the same tracking
>> infrastructure for (a). In other words, it should not be necessary to
>> add the specific fields "reboot_timeout" and "boot_splash" to FWCfgState.
>
> OK, I'll drop these patches from this series.
>
>>
>> Thanks,
>> Laszlo
>>
>>>
>>> diff --git a/hw/nvram/fw_cfg.c b/hw/nvram/fw_cfg.c
>>> index b73a591eff..182d27f59a 100644
>>> --- a/hw/nvram/fw_cfg.c
>>> +++ b/hw/nvram/fw_cfg.c
>>> @@ -250,7 +250,9 @@ static void fw_cfg_reboot(FWCfgState *s)
>>> }
>>> }
>>>
>>> - fw_cfg_add_file(s, "etc/boot-fail-wait", g_memdup(&rt_val, 4), 4);
>>> + s->reboot_timeout = rt_val;
>>> + fw_cfg_add_file(s, "etc/boot-fail-wait",
>>> + &s->reboot_timeout, sizeof(s->reboot_timeout));
>>> }
>>>
>>> static void fw_cfg_write(FWCfgState *s, uint8_t value)
>>> diff --git a/include/hw/nvram/fw_cfg.h b/include/hw/nvram/fw_cfg.h
>>> index 828ad9dedc..99f6fafcaa 100644
>>> --- a/include/hw/nvram/fw_cfg.h
>>> +++ b/include/hw/nvram/fw_cfg.h
>>> @@ -53,6 +53,8 @@ struct FWCfgState {
>>> dma_addr_t dma_addr;
>>> AddressSpace *dma_as;
>>> MemoryRegion dma_iomem;
>>> +
>>> + uint32_t reboot_timeout;
>>> };
>>>
>>> struct FWCfgIoState {
>>>
>>
