Re: [GIT] Networking

From: Marc Zyngier
Date: Tue Jan 19 2016 - 03:55:46 EST


On 18/01/16 22:05, Arend van Spriel wrote:
> On 18-1-2016 12:30, Marc Zyngier wrote:
>> Hi Kalle,
>>
>> On 16/01/16 11:57, Kalle Valo wrote:
>>> Marc Zyngier <marc.zyngier@xxxxxxx> writes:
>>>
>>>> David, Hante,
>>>>
>>>> On 13/01/16 02:51, David Miller wrote:
>>>>
>>>> [...]
>>>>
>>>>> Hante Meuleman (33):
>>>> [...]
>>>>> brcmfmac: Move all module parameters to one place
>>>
>>> As a reminder to myself this is the commit id:
>>>
>>> 7d34b0560567 brcmfmac: Move all module parameters to one place
>>>
>>>> This particular patch breaks one of my boxes in a spectacular way:
>>>>
>>>> [ 3.602155] Unable to handle kernel paging request at virtual address 000027e4
>>>> [ 3.602160] pgd = c0003000
>>>> [ 3.602169] [000027e4] *pgd=80000040004003, *pmd=00000000
>>>> [ 3.602181] Internal error: Oops: 206 [#1] PREEMPT SMP ARM
>>>
>>> [...]
>>>
>>>> This is caused by this hunk:
>>>>
>>>> @@ -890,7 +887,8 @@ static void brcmf_sdiod_sgtable_alloc(struct brcmf_sdio_dev *sdiodev)
>>>> if (!sdiodev->sg_support)
>>>> return;
>>>>
>>>> - nents = max_t(uint, BRCMF_DEFAULT_RXGLOM_SIZE, brcmf_sdiod_txglomsz);
>>>> + nents = max_t(uint, BRCMF_DEFAULT_RXGLOM_SIZE,
>>>> + sdiodev->bus_if->drvr->settings->sdiod_txglomsz);
>>>> nents += (nents >> 4) + 1;
>>>>
>>>> WARN_ON(nents > sdiodev->max_segment_count);
>>>>
>>>> were drvr->settings is NULL (as the settings allocation seems to be done
>>>> much later). The fix is not completely obvious to me (probably requires
>>>> pushing the call to brcmf_mp_device_attach() down into the various bus
>>>> specific functions). An alternative would be to restore the txglomsz
>>>> parameter as it was before and not rely on settings being allocated.
>>>
>>> Should we revert the patch or can you Hante fix this? The revert doesn't
>>> seem to be trivial so I would appreciate if someone could send a patch.
>>
>> I've worked out a partial revert (see below) that allows my system to
>> boot, but I'd rather see a proper fix from the maintainer of this code.
>
> Hi Marc,
>
> Thanks for the patch, but Hante has created a different patch basically
> deferring the allocation of the sgtable. Feel free to give it a spin on
> your box and share the results.

Hi Arend,

This patch fixes indeed the problem, thanks (I had to undo the mangling
your mailer had done, though). So feel free to add my:

Reported-by: Marc Zyngier <marc.zyngier@xxxxxxx>
Tested-by: Marc Zyngier <marc.zyngier@xxxxxxx>

It would be good if this could make it in -rc1.

> I am looking for a arm64 platform to add to my test machines so if
> you can recommend one (with PCIe).

Note that this crash happened with 32bit ARM, without PCIe (the box is a
"Cubietruck").

As for arm64, there is now quite a few machines out there (ARM sells a
dev board called Juno, but there is also various offerings with AMD
Opteron A1100, Cavium ThunderX, APM X-Gene, and maybe even some Broadcom
based systems, who knows! ;-).

Thanks,

M.
--
Jazz is not dead. It just smells funny...