Bug in drivers/net/ethernet/freescale/fec_main.c, TX is broken. In 4.0.0-rc3

[Панов Андрей]

Hi!

16.03.2015, 17:01, "fugang.duan at freescale.com" <fugang.duan at freescale.com>:
>>>
>>>  Do you have any lost for reproduce the issue ?
>>  Current net tree does not have this issue, it works fine. Thanks to all.
>>
>>  From git://git.kernel.org/pub/scm/linux/kernel/git/davem/net.git
>>
>>  apx at astra:~$ uname -r
>>  4.0.0-rc3-00150-g10640d3-dirty
>>
>
> I test the commit f00bbd2 that include the patch 2b995f63987.
> But cannot reproduce your issue. Pls double confirm it in your board.

Confirmed that net-next is buggy.

Client kernel version:
apx at astra:~$ uname -r
4.0.0-rc3-00875-gf00bbd2-dirty

100M file of zeros on client:
apx at astra:~$ ls -l zeros 
-rw-r--r-- 1 apx apx 104857600 мар 10 22:33 zeros

File contents (ran on client host):
apx at astra:~$ hd zeros 
00000000  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |................|
*
06400000

NFS server kernel version:
apx at ct:~$ uname -r
4.0.0-rc2-00480-g29e70e6

100M file of zeros on client:
apx at ct:~$ ls -l zeros 
-rw-r--r-- 1 apx apx 104857600 мар 10 22:33 zeros

File contents (ran on server host):
apx at ct:~$ hd zeros    
00000000  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |................|
*
001d4a80  1c f0 9f e5 1c f0 9f e5  1c f0 9f e5 1c f0 9f e5  |................|
*
001d4aa0  1c f0 9f e5 74 fb 00 00  bc ff 93 00 c0 ff 93 00  |....t...........|
001d4ab0  c4 ff 93 00 c8 ff 93 00  cc ff 93 00 d0 ff 93 00  |................|
001d4ac0  d4 ff 93 00 d8 ff 93 00  13 00 00 00 28 63 29 20  |............(c) |
001d4ad0  43 6f 70 79 72 69 67 68  74 20 32 30 30 37 2d 32  |Copyright 2007-2|
001d4ae0  30 31 32 2c 20 46 72 65  65 73 63 61 6c 65 20 53  |012, Freescale S|
001d4af0  65 6d 69 63 6f 6e 64 75  63 74 6f 72 2e 20 41 6c  |emiconductor. Al|
001d4b00  6c 20 72 69 67 68 74 73  20 72 65 73 65 72 76 65  |l rights reserve|
001d4b10  64 2e 00 00 dd 00 2c 41  11 73 00 00 d3 74 00 00  |d.....,A.s...t..|
001d4b20  3d 75 00 00 a9 78 00 00  4f 78 00 00 75 77 00 00  |=u...x..Ox..uw..|
001d4b30  07 76 00 00 c3 79 00 00  09 7a 00 00 75 7a 00 00  |.v...y...z..uz..|

Any other client works fine with this server.
And note that corruption is not random, it looks like a zero page (vectors at beginning),
so somewhere pointer to data is corrupted (and set to zero(?)).
This is repeated many times.

I use an Embedsky E9 board: http://en.embedsky.com/product_info.php?cateid=169&id=169
It is SabreSD-alike board with a fewer peripherials, and at network side it has a different PHY,
instead of Atheros it has a Realtek RTL8211E, and I have to initialize it this way (board early fixup patch):

--- a/arch/arm/mach-imx/mach-imx6q.c
+++ b/arch/arm/mach-imx/mach-imx6q.c
@@ -166,6 +166,19 @@ static int ar8035_phy_fixup(struct phy_device *dev)
 
 #define PHY_ID_AR8035 0x004dd072
 
+static int rtl8211e_phy_fixup(struct phy_device *dev)
+{
+       phy_write(dev, 0x00, 0x3140);
+       msleep(10);
+       phy_write(dev, 0x00, 0x3340);
+       msleep(10);
+
+       return 0;
+}
+
+#define PHY_ID_RTL8211E 0x001cc915
+#define REALTEK_PHY_ID_MASK 0x001fffff
+
 static void __init imx6q_enet_phy_init(void)
 {
        if (IS_BUILTIN(CONFIG_PHYLIB)) {
@@ -177,6 +190,8 @@ static void __init imx6q_enet_phy_init(void)
                                ar8031_phy_fixup);
                phy_register_fixup_for_uid(PHY_ID_AR8035, 0xffffffef,
                                ar8035_phy_fixup);
+               phy_register_fixup_for_uid(PHY_ID_RTL8211E, REALTEK_PHY_ID_MASK,
+                               rtl8211e_phy_fixup);
        }
 }

(LAKML folks told me this should go somewhere in network driver, but now it is there)

Without this there is no network at all, and I had ran this code for year without any glitch. 3.19 kernel works fine.
(It causes "-dirty" in kernel version)
And different phy initialization cannot cause a non-random stream corruption.

I've found commit that produces a bug by looking a changes between 3.19 (surely working)
 and 4.0.0-rc3 (surely non-working) in freescale network driver.

Hope this helps.

--
 Андрей