Re: [PATCH] mips: bmips: BCM6358: disable arch_sync_dma_for_cpu_all()

[Florian Fainelli <f.fainelli@gmail.com>]

On 3/11/2023 11:31 PM, William Zhang wrote:
> 
> 
> On 03/11/2023 11:44 AM, Jonas Gorski wrote:
>> On Sat, 11 Mar 2023 at 18:32, Florian Fainelli <f.fainelli@gmail.com> 
>> wrote:
>>>
>>>
>>>
>>> On 3/10/2023 4:13 AM, Álvaro Fernández Rojas wrote:
>>>> arch_sync_dma_for_cpu_all() causes kernel panics on BCM6358 with 
>>>> EHCI/OHCI:
>>>> [    3.881739] usb 1-1: new high-speed USB device number 2 using 
>>>> ehci-platform
>>>> [    3.895011] Reserved instruction in kernel code[#1]:
>>>> [    3.900113] CPU: 0 PID: 1 Comm: init Not tainted 5.10.16 #0
>>>> [    3.905829] $ 0   : 00000000 10008700 00000000 77d94060
>>>> [    3.911238] $ 4   : 7fd1f088 00000000 81431cac 81431ca0
>>>> [    3.916641] $ 8   : 00000000 ffffefff 8075cd34 00000000
>>>> [    3.922043] $12   : 806f8d40 f3e812b7 00000000 000d9aaa
>>>> [    3.927446] $16   : 7fd1f068 7fd1f080 7ff559b8 81428470
>>>> [    3.932848] $20   : 00000000 00000000 55590000 77d70000
>>>> [    3.938251] $24   : 00000018 00000010
>>>> [    3.943655] $28   : 81430000 81431e60 81431f28 800157fc
>>>> [    3.949058] Hi    : 00000000
>>>> [    3.952013] Lo    : 00000000
>>>> [    3.955019] epc   : 80015808 setup_sigcontext+0x54/0x24c
>>>> [    3.960464] ra    : 800157fc setup_sigcontext+0x48/0x24c
>>>> [    3.965913] Status: 10008703       KERNEL EXL IE
>>>> [    3.970216] Cause : 00800028 (ExcCode 0a)
>>>> [    3.974340] PrId  : 0002a010 (Broadcom BMIPS4350)
>>>> [    3.979170] Modules linked in: ohci_platform ohci_hcd 
>>>> fsl_mph_dr_of ehci_platform ehci_fsl ehci_hcd gpio_button_hotplug 
>>>> usbcore nls_base usb_common
>>>> [    3.992907] Process init (pid: 1, threadinfo=(ptrval), 
>>>> task=(ptrval), tls=77e22ec8)
>>>> [    4.000776] Stack : 81431ef4 7fd1f080 81431f28 81428470 7fd1f068 
>>>> 81431edc 7ff559b8 81428470
>>>> [    4.009467]         81431f28 7fd1f080 55590000 77d70000 77d5498c 
>>>> 80015c70 806f0000 8063ae74
>>>> [    4.018149]         08100002 81431f28 0000000a 08100002 81431f28 
>>>> 0000000a 77d6b418 00000003
>>>> [    4.026831]         ffffffff 80016414 80080734 81431ecc 81431ecc 
>>>> 00000001 00000000 04000000
>>>> [    4.035512]         77d54874 00000000 00000000 00000000 00000000 
>>>> 00000012 00000002 00000000
>>>> [    4.044196]         ...
>>>> [    4.046706] Call Trace:
>>>> [    4.049238] [<80015808>] setup_sigcontext+0x54/0x24c
>>>> [    4.054356] [<80015c70>] setup_frame+0xdc/0x124
>>>> [    4.059015] [<80016414>] do_notify_resume+0x1dc/0x288
>>>> [    4.064207] [<80011b50>] work_notifysig+0x10/0x18
>>>> [    4.069036]
>>>> [    4.070538] Code: 8fc300b4  00001025  26240008 <ac820000> 
>>>> ac830004  3c048063  0c0228aa  24846a00  26240010
>>>> [    4.080686]
>>>> [    4.082517] ---[ end trace 22a8edb41f5f983b ]---
>>>> [    4.087374] Kernel panic - not syncing: Fatal exception
>>>> [    4.092753] Rebooting in 1 seconds..
>>>
>>> Did you pinpoint which specific instruction within
>>> arch_sync_dma_for_cpu_all() is causing the reserved instruction 
>>> exception?
>>
>> It's setup_sigcontext(), not arch_sync_dma_for_cpu_all() that's
>> causing the exception ;-)
>>
>> Hand decoding the Code gives me
>>
>> lw $1, 0xb4($fp)
>> or $v0, 0, 0
>> addiu $a0, $s1, 8
>> sw $v0, 0($a0) <- the code in brackets, so I guess EPC?
>> sw $v1, 4($a0)
>>
>> which I assume is this part:
>>
>> err |= __put_user(regs->cp0_epc, &sc->sc_pc);
>>
>> (0xb4 is the offset of cp0_epc, 0x8 the offset of sc_pc)
>>
>> One thing I see is that we do the RAC flush for BMIPS3300, 4350 and
>> 4380, but only initialize it for 3300 [1], but leave it at whatever
>> state the bootloader did for the other ones. Maybe it has some invalid
>> config in (that particuar?) 6358 that triggers issues later on after a
>> flush? E.g. the flush puts it in an error state, and the next time
>> something triggers a prefetch(write?) (by trying to access userspace)
>> it generates an error exception.
>>
> Depending on the bootloader but likely bootloader does not use RAC at 
> all.  So agree that RAC may not be properly initialized when the flush 
> function is called and push the stale data to corrupt memory and cause 
> problem later on the userspace.

Alvaro, could you do the following and let us know the results, at boot 
time in bcm6358_quirks():

- issue a read_c0_brcm_config_0() and look whether bit 29 (RAC present) 
is set, if it is not set, then we should forcibly disable the use of the 
RAC using a flag

- issue a __raw_readl(cbr + BMIPS_RAC_CONFIG) and check whether bits 0 
(RAC_I) or 1 (RAC_D) are set, if not, try to set them and see whether it 
works

Thanks!
-- 
Florian