On Tue, Jun 18, 2013 at 12:04:51PM +0800, Hanjun Guo wrote:
On 2013-6-15 15:55, Hanjun Guo wrote:
On 2013-6-14 21:17, Tomasz Nowicki wrote:
W dniu 14.06.2013 15:08, Hanjun Guo pisze:
On 2013-6-14 20:23, Tomasz Nowicki wrote:
W dniu 14.06.2013 11:53, Hanjun Guo pisze:
PM Event and Control registers and other ACPI fixed hardware features were not supported on ARM64 yet, so we need enable reduced hardware as described in ACPI 5.0.
I met a boot panic failure on the ARMv8 model,
[ 9.024370] Unable to handle kernel paging request at virtual address ffffffbffbe00001 [ 9.024423] pgd = ffffffc00007d000 [ 9.024468] [ffffffbffbe00001] *pgd=000000008007f003, *pmd=0000000000000000 [ 9.024538] Internal error: Oops: 96000006 [#1] SMP [ 9.024577] Modules linked in: [ 9.024627] CPU: 0 Not tainted (3.9.0+ #3) [ 9.024691] PC is at acpi_os_read_port+0x58/0xc8 [ 9.024753] LR is at acpi_hw_read_port+0x4c/0xc4 [ 9.024810] pc : [<ffffffc0002810cc>] lr : [<ffffffc0002a4fc4>] pstate: 600003c5 .... [ 9.031956] Call trace: [ 9.032021] [<ffffffc0002810cc>] acpi_os_read_port+0x58/0xc8 [ 9.032094] [<ffffffc0002a4fc4>] acpi_hw_read_port+0x4c/0xc4 [ 9.032165] [<ffffffc0002a4198>] acpi_hw_read+0x6c/0x100 [ 9.032237] [<ffffffc0002a4250>] acpi_hw_read_multiple+0x24/0x70 [ 9.032312] [<ffffffc0002a457c>] acpi_hw_register_read+0xa8/0x164 [ 9.032386] [<ffffffc0002a52dc>] acpi_write_bit_register+0x9c/0x194 [ 9.032472] [<ffffffc0002bffb4>] acpi_processor_get_power_info+0x6c4/0x748 [ 9.032550] [<ffffffc000565220>] acpi_processor_power_init+0xac/0x138 [ 9.032630] [<ffffffc0003f3c00>] acpi_processor_start+0x48/0x138 [ 9.032704] [<ffffffc000565118>] acpi_processor_add+0x43c/0x498 [ 9.032784] [<ffffffc000283a74>] acpi_device_probe+0x3c/0x198 [ 9.032862] [<ffffffc0002ef0fc>] driver_probe_device+0x90/0x348 [ 9.032940] [<ffffffc0002ef450>] __driver_attach+0x9c/0xa0 [ 9.033015] [<ffffffc0002ed40c>] bus_for_each_dev+0x4c/0x8c [ 9.033090] [<ffffffc0002eeb98>] driver_attach+0x20/0x28 [ 9.033166] [<ffffffc0002ee6e0>] bus_add_driver+0xfc/0x254 [ 9.033244] [<ffffffc0002ef8b0>] driver_register+0x6c/0x184 [ 9.033325] [<ffffffc000284894>] acpi_bus_register_driver+0x34/0x44 [ 9.033400] [<ffffffc000559c60>] acpi_processor_init+0x28/0x40 [ 9.033469] [<ffffffc000081438>] do_one_initcall+0x100/0x138 [ 9.033544] [<ffffffc0005448c8>] kernel_init_freeable+0x128/0x1cc [ 9.033620] [<ffffffc0003f2f90>] kernel_init+0x10/0xcc [ 9.033700] Code: d2bf7c02 f2dff7e2 f2ffffe2 8b020000 (79400000) [ 9.033843] ---[ end trace 64376967e6bc20a9 ]--- [ 9.033995] Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
Since PM control register is not exist on ARM64, It will cause such failure.
Does PM control register exist on ARM?
PM control registers here are ACPI PM control registers, which described in ACPI5.0 section 4.8, so I think they are not exist on ARM at present.
How it is possible it didn't crash on Arndale? I suppose we have the same functionality for Arndale as for Foundation model.
Good question. To tell the truth, I not sure why it is not happened on Arndale, and even on Graeme's ARMv8 foundation model, we have the same version of foundation model, the same kernel code and the config file, but it not happened on Graeme's model. (odd, right? )
I guess on my armv8 model, we lucky hit the bug that it access the virtual address which not exist, but on Graeme's model, the virtual address is available.
I'm not sure of this, I will ask some guy who is a expert on memory management next Monday.
Ok, I have tracked this problem down, and I will explain why this is not happened on Arndale and armv8 model with more than one core booted.
Let's go back to drivers/acpi/processor_driver.c, acpi_processor_start() will be called in acpi_processor_add() when need_hotplug_init flag is 0, when the system booted, the need_hotplug_init flag should be 0, but on the system with multi cpu booted on arm:
acpi_processor_add()
acpi_processor_get_info()
acpi_get_cpuid() if cpu is hot-added, will return -1 but on ARM platform, map gic id to cpu id is not implemented, so acpi_get_cpuid() will always return -1 on arm. acpi_processor_hotadd_init() will be called, and set need_hotplug_init to 1acpi_processor_start() will not called, and the bug will not triggered.
On the UP sytem: acpi_processor_hotadd_init() will not be called, so the need_hotplug_init flag is zero. and acpi_processor_start() will called with bug happened.
Arndale board has two cpu cores, so this is not happened on Arndale, if you want to trigger this bug, boot the acpi kernel on armv8 model with --cores=1.
I will fix this in cpu topology cards, and enhance hardware reduced for acpi driver is also needed.
Hi Hanjun,
excellent peice of investigation, I never even considered to ask how many cores people were starting with. I had --cores=4.
Thanks
Graeme