On 12/30/2014 03:05 AM, Tomasz Figa wrote:
> Thanks a lot for investigating this, even before I could look into
> splitting this.
>
> 2014-12-30 3:23 GMT+09:00 Nishanth Menon <nm@xxxxxx>:
>> On 12/23/2014 04:48 AM, Marek Szyprowski wrote:
>>
>>> -static void l2c310_resume(void)
>>> +static void l2c310_configure(void __iomem *base)
>>> {
>>> - void __iomem *base = l2x0_base;
>>> + unsigned revision;
>>>
>>> - if (!(readl_relaxed(base + L2X0_CTRL) & L2X0_CTRL_EN)) {
>>> - unsigned revision;
>>> -
>>> - /* restore pl310 setup */
>>> - writel_relaxed(l2x0_saved_regs.tag_latency,
>>> - base + L310_TAG_LATENCY_CTRL);
>>> - writel_relaxed(l2x0_saved_regs.data_latency,
>>> - base + L310_DATA_LATENCY_CTRL);
>>> - writel_relaxed(l2x0_saved_regs.filter_end,
>>> - base + L310_ADDR_FILTER_END);
>>> - writel_relaxed(l2x0_saved_regs.filter_start,
>>> - base + L310_ADDR_FILTER_START);
>>> -
>>> - revision = readl_relaxed(base + L2X0_CACHE_ID) &
>>> - L2X0_CACHE_ID_RTL_MASK;
>>> -
>>> - if (revision >= L310_CACHE_ID_RTL_R2P0)
>>> - l2c_write_sec(l2x0_saved_regs.prefetch_ctrl, base,
>>> - L310_PREFETCH_CTRL);
>>> - if (revision >= L310_CACHE_ID_RTL_R3P0)
>>> - l2c_write_sec(l2x0_saved_regs.pwr_ctrl, base,
>>> - L310_POWER_CTRL);
>>> -
>>> - l2c_enable(base, l2x0_saved_regs.aux_ctrl, 8);
>>> -
>>> - /* Re-enable full-line-of-zeros for Cortex-A9 */
>>> - if (l2x0_saved_regs.aux_ctrl & L310_AUX_CTRL_FULL_LINE_ZERO)
>>> - set_auxcr(get_auxcr() | BIT(3) | BIT(2) | BIT(1));
>>> - }
>>> + /* restore pl310 setup */
>>> + writel_relaxed(l2x0_saved_regs.tag_latency,
>>> + base + L310_TAG_LATENCY_CTRL);
>>> + writel_relaxed(l2x0_saved_regs.data_latency,
>>> + base + L310_DATA_LATENCY_CTRL);
>>> + writel_relaxed(l2x0_saved_regs.filter_end,
>>> + base + L310_ADDR_FILTER_END);
>>> + writel_relaxed(l2x0_saved_regs.filter_start,
>>> + base + L310_ADDR_FILTER_START);
>>> +
>>
>> ^^ The above change broke AM437xx. Looks like the change causes the
>> following behavior difference on AM437x. For some reason, touching any
>> of the above 4 registers(even with the values read from the same
>> registers) causes AM437x to go beserk. Comment the 4 writes and we
>> reach shell. looks like l2c310_resume is not invoked prior to this
>> series. :(.. now that we reuse that logic to actually do programming,
>> we start to see the problem.
>
> Hmm, but the thing is that .configure() should not be called if the
> controller is already configured, i.e. L2X0_CTRL_EN in L2X0_CTRL is
> set. Maybe I missed some check somewhere. Let me reread my code I
> wrote quite a long time ago and make sure.
you have'nt missed a check here. it does indeed get called
l2c310_enable->l2c_enable -> (if cache
disabled)->l2c_configure->l2c310_configure
It looks like a quirk of AM437x which remained hidden till this patch
exposed it. The original pl310 resume would have been invoked if
outer_resume was invoked (if my reading of code is correct), which in
the case of AM437x was never invoked during boot. By reusing the
restore code for resume (which in my opinion is a good change), we
seemed to have exposed quirky behavior on am437x. I started a thread
yesterday with hardware folks trying to understand the integration
aspects and explanation for this quirky behavior - unfortunately, with
holidays, I doubt I might get a quick answer fast. but the workaround
seems obvious - do not write to the the mentioned 4 registers on
am437x. This might make sense it arm,tag-latency , arm,data-latency,
arm,filter-ranges properties are not defined in OF.
>> one option might be to write only those registers that differ from
>> saved_registers (example: unmodified values dont need reprogramming).
>>
>> Looks like the following also need addressing:
>> data->save is called twice (once more after l2cof_init)
>> l2c310_init_fns also needs l2c310_configure
>> will be nice to use l2x0_data only after we kmemdup data in __l2c_init
>
> I'll check this.
Thanks.
>
>>
>> if you'd like to split this up in pieces, [1] might be nice - will go
>> good to change the pl310, aurora etc in each chunks to enable better
>> review.
>
> Thanks a lot, the split up version will be definitely useful. Just to
> make sure, the parts look quite bisectable, but have you verified that
> applying the changes one by one leave the L2 cache working on OMAP?
The split is not complete or properly done as I ignored aurora and was
trying to do it focussed on pl310 impacted code in a hurry, but yes,
the split is indeed bisectable for OMAP platforms. The intent was to
indicate the direction we should probably take and introduce the
refactor in stages. At the very least, it tends to be easier to debug
down to.
>> [1]
>> https://github.com/nmenon/linux-2.6-playground/commits/temp/l2c-patch2-splitup
>
> Best regards,
> Tomasz
>
--
Regards,
Nishanth Menon
--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
