Hi,
On 4/7/20 10:23 AM, Ahmad Fatoum wrote:
> Hello Giorgio,
>
> On 4/7/20 9:46 AM, Giorgio wrote:
>> Hi,
>>
>> OK, my current fix follows also the same logic as for the values
>> of 'vbar' and 'ttb' in armv7_secure_monitor_install(). I've also
>> found the 'set_domain(v)' in mmu.h, don't need to write a new one.
>>
>> I'll send a patch for this.
>
> Great. Even better than hardcoding the CLIENT_DOMAIN.
>
OK.
To read the current value of DACR, in secure mode, we need a
get_domain(). I would add it to mmu.h, beside the set_domain().
>> What do you mean with the 'other i.MX7 patches' ?
>
> Didn't you add support for some i.MX7 spi flash image format?
>
That was an evil hack actually, just to verify why the normal barebox image
didn't worked. To really support booting barebox from the qspi flash
I think we need more *structural* changes to the way barebox starts.
For this I need *at least* some suggestions from someone that really knows
in detail how barebox works and how an image is built, like you or Sascha...
Maybe what can be committed is the dts block in imx7s.dtsi for the qspi that
enables using the flash as normal runtime mass storage. This should be not
so dangerous because the qspi becomes only active when its status is also enabled.
>> I've also noticed that some asm() statements in sm.c lack the 'volatile'
>> attribute.
>
> Are they strictly necessary? My understanding is that we need
> asm volatile on no output operands, but side effects.
>
> All asm statements in the file either lack side effects
> (the reads) or have no output operands (the writes).
>
> That said, I don't mind making them all volatile just
> to be a little extra sure..
>
I think for the current code in sm.c we don't strictly need the volatiles,
I noticed problems in debug code like this:
static u32 read_dacr(void)
{
unsigned int reg;
// without volatile:
asm ("mrc p15, 0, %0, c3, c0, 0\n" : "=r"(reg));
return reg;
}
...
printf("%s: 1) dacr: 0x%08x\n", __func__, read_dacr());
/* Initialize the secure monitor */
arm_smccc_smc(0, 0, 0, 0, 0, 0, 0, 0, &res);
/* We're in nonsecure mode now */
printf("%s: 2) dacr: 0x%08x\n", __func__, read_dacr());
Here the compiler could think he can call read_dacr() only once,
cache the result in a register and use its value two times. I verified
that for the previous debug code the volatile attribute makes a
difference.
giorgio
> Cheers,
> Ahmad
>
>>
>> giorgio
>>
>> On 4/6/20 8:44 PM, Ahmad Fatoum wrote:
>>> Hello Giorgio,
>>>
>>> On 4/6/20 5:15 PM, Giorgio wrote:
>>>> Searching in the Armv7-A ref. man. here:
>>>>
>>>> https://static.docs.arm.com/ddi0406/cd/DDI0406C_d_armv7ar_arm.pdf
>>>>
>>>> at B4.1.43 (page B4 - 1554) I found that the reset value of this cp15 reg.
>>>> is UNKNOWN and I can also verify this. Moreover the reg. must also be
>>>> banked: there is a copy in secure mode and one in nonsecure mode.
>>>> My test consists in reading the register just before and just after
>>>> the call to arm_smccc_smc(0, 0, 0, 0, 0, 0, 0, 0, &res):
>>>>
>>>> static u32 read_dacr(void)
>>>> {
>>>> unsigned int reg;
>>>>
>>>> asm volatile ("mrc p15, 0, %0, c3, c0, 0\n" : "=r"(reg));
>>>> return reg;
>>>> }
>>>>
>>>> ...
>>>>
>>>> int __armv7_secure_monitor_install(void)
>>>> {
>>>> ...
>>>> printf("%s: 1) dacr: 0x%08x\n",__func__,read_dacr());
>>>>
>>>> /* Initialize the secure monitor */
>>>> arm_smccc_smc(0, 0, 0, 0, 0, 0, 0, 0, &res);
>>>>
>>>> /* We're in nonsecure mode now */
>>>>
>>>> printf("%s: 2) dacr: 0x%08x\n",__func__,read_dacr());
>>>>
>>>> return 0;
>>>> }
>>>>
>>>> and here is what I get:
>>>>
>>>> samx7: / smc -n
>>>> __armv7_secure_monitor_install: 1) dacr: 0x00000001
>>>> __armv7_secure_monitor_install: 2) dacr: 0xdbf7afaa
>>>>
>>>> If I write a 0x1 or a 0x3 to DACR just before enabling the MMU in
>>>> nonsecure mode then barebox does not hang.
>>>
>>> Oh! seems i have been lucky so far:
>>>
>>> __armv7_secure_monitor_install: 1) dacr: 0x00000001
>>> __armv7_secure_monitor_install: 2) dacr: 0xa133ad17
>>>
>>> Yes, we will want to call set_domain(DOMAIN_CLIENT)
>>> somewhere, maybe just before set_ttbr?
>>>
>>> Do you want to send a patch?
>>> (Your other i.MX7 patches are welcome as well :-)
>>>
>>> Cheers and thanks for debugging this
>>> Ahmad
>>>
>>>>
>>>> giorgio
>>>>
>>>> On 2020-04-06 08:16, Ahmad Fatoum wrote:
>>>>> Hello Giorgio,
>>>>>
>>>>> On 4/3/20 3:47 PM, Giorgio wrote:
>>>>>> Hi Ahmad,
>>>>>>
>>>>>> thank you for the detailed explanations, I'll have a look
>>>>>> at the armv7 ref. manual for more background.
>>>>>>
>>>>>> I wanted just to note, the problem is specifically linked
>>>>>> to enabling the MMU:
>>>>>>
>>>>>> in arch/arm/cpu/cache-armv7.S:
>>>>>>
>>>>>>
>>>>>> orrne r0, r0, #1 @ MMU enabled
>>>>>> ...
>>>>>> mcr p15, 0, r0, c1, c0, 0 @ load control register
>>>>>>
>>>>>> without the 'orrne ...' the imx7 does not hang.
>>>>>
>>>>> I can't reproduce this exact problem. My setup:
>>>>>
>>>>> - i.MX7D sabresd board,
>>>>> - imx_v7_defconfig
>>>>> - check out of upstream/next, commit 5931fe40 ("Merge branch 'for-next/zii' into next")
>>>>> - revert of commit 8b2104d ("driver: Call of_clk_set_defaults for each probed device")
>>>>> - nv bootm.secure_state=nonsecure
>>>>>
>>>>> With this, I didn't observe any barebox hangs[1] while preparing a Linux net boot.
>>>>> What's your setup?
>>>>>
>>>>> [1]: Linux still hangs due to what I assume to be a psci issue, kernel log says
>>>>> "unsupported enable-method property: psci" before getting stuck durcing SDHCI probe
>>>>>
>>>>
>>>
>>
>
_______________________________________________
barebox mailing list
barebox@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/barebox
