[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH] hw/char/exynos4210_uart: Fix memleaks in exynos4210_uart_ini

From: Philippe Mathieu-Daudé
Subject: Re: [PATCH] hw/char/exynos4210_uart: Fix memleaks in exynos4210_uart_init
Date: Thu, 13 Feb 2020 17:32:57 +0100
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.4.1

On 2/13/20 3:28 PM, Markus Armbruster wrote:
Eduardo Habkost <address@hidden> writes:

On Wed, Feb 12, 2020 at 08:39:55AM +0100, Philippe Mathieu-Daudé wrote:
Cc'ing Eduardo & Markus.

On 2/12/20 7:44 AM, Chenqun (kuhn) wrote:
-----Original Message-----
From: Philippe Mathieu-Daudé [mailto:address@hidden]
Sent: Wednesday, February 12, 2020 2:16 PM
To: Chenqun (kuhn) <address@hidden>; qemu-
address@hidden; address@hidden; address@hidden
Cc: address@hidden; Zhanghailiang
Subject: Re: [PATCH] hw/char/exynos4210_uart: Fix memleaks in

On 2/12/20 4:36 AM, address@hidden wrote:
From: Chen Qun <address@hidden>

It's easy to reproduce as follow:
virsh qemu-monitor-command vm1 --pretty '{"execute":
"device-list-properties", "arguments":{"typename":"exynos4210.uart"}}'

ASAN shows memory leak stack:
     #1 0xfffd896d71cb in g_malloc0 (/lib64/libglib-2.0.so.0+0x571cb)
     #2 0xaaad270beee3 in timer_new_full /qemu/include/qemu/timer.h:530
     #3 0xaaad270beee3 in timer_new /qemu/include/qemu/timer.h:551
     #4 0xaaad270beee3 in timer_new_ns /qemu/include/qemu/timer.h:569
     #5 0xaaad270beee3 in exynos4210_uart_init
     #6 0xaaad275c8f4f in object_initialize_with_type /qemu/qom/object.c:516
     #7 0xaaad275c91bb in object_new_with_type /qemu/qom/object.c:684
     #8 0xaaad2755df2f in qmp_device_list_properties

Reported-by: Euler Robot <address@hidden>
Signed-off-by: Chen Qun <address@hidden>
    hw/char/exynos4210_uart.c | 8 ++++----
    1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/hw/char/exynos4210_uart.c b/hw/char/exynos4210_uart.c
index 25d6588e41..5048db5410 100644
--- a/hw/char/exynos4210_uart.c
+++ b/hw/char/exynos4210_uart.c
@@ -674,10 +674,6 @@ static void exynos4210_uart_init(Object *obj)
        SysBusDevice *dev = SYS_BUS_DEVICE(obj);
        Exynos4210UartState *s = EXYNOS4210_UART(dev);

-    s->fifo_timeout_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
-                                         exynos4210_uart_timeout_int, s);
-    s->wordtime = NANOSECONDS_PER_SECOND * 10 / 9600;

Why are you moving s->wordtime from init() to realize()?

Hi  Philippe,  thanks for your reply!

Because I found the variable wordtime is usually used with fifo_timeout_timer.
Eg, they are used together in the exynos4210_uart_rx_timeout_set function.

I didn't find anything wrong with wordtime in the realize().
Does it have any other effects?

IIUC when we use both init() and realize(), realize() should only contains
on code that consumes the object properties... But maybe the design is not
clear. Then why not move all the init() code to realize()?

Normally I would recommend the opposite: delay as much as
possible to realize(), to avoid unwanted side effects when (e.g.)
running qom-list-properties.

Sadly, our documentation on device initialization and realization is
rather sparse, and developers are left guessing.  Their guesses are
often based on what existing code does.  Some of the existing code even
gets things right.

A few rules from the top of my head:

Worth a new thread...

* Creating and immediately destroying an object must be safe and free of
   side effects: initialization may only touch the object itself, and
   finalization must clean up everything initialization creates.

* unrealize() must clean up everything realize() creates.

Hmm I guess remember someone once said "only for hot-pluggable objects, else don't bother". But then we make a non-hot-pluggable object as hot-pluggable and have to fix leaks. Or we start a new hot-pluggable device based on some code without unrealize()...

* Since initialization cannot fail, code that needs to fail gracefully
   must live in realize().

* Since property values get set between initialization and realization,
   code that uses property values must live in realize().

* Dynamic properties have to be created in initialization to be visible
   in introspection.

But as s->wordtime is a simple struct field (that we could even
decide to expose to the outside as a read-only QOM property), it
doesn't really matter.  Personally, I would keep it where it is
just to avoid churn.

reply via email to

[Prev in Thread] Current Thread [Next in Thread]