Re: Bug report: kernel paniced while booting

From: Song Shuai
Date: Tue Jun 06 2023 - 06:05:18 EST



在 2023/6/5 22:25, Alexandre Ghiti 写道:
Hi Song,

On Mon, Jun 5, 2023 at 12:52 PM Song Shuai <songshuaishuai@xxxxxxxxxxx> wrote:

Description of problem:

Booting Linux With RiscVVirtQemu edk2 firmware, a Store/AMO page fault was trapped to trigger a kernel panic.
The entire log has been posted at this link : https://termbin.com/nga4.

You can reproduce it with the following step :

1. prepare the environment with
- Qemu-virt: v8.0.0 (with OpenSbi v1.2)
- edk2 : at commit (2bc8545883 "UefiCpuPkg/CpuPageTableLib: Reduce the number of random tests")
- Linux : v6.4-rc1 and later version

2. start the Qemu virt board

```sh
$ cat ~/8_riscv/start_latest.sh
#!/bin/bash
/home/song/8_riscv/3_acpi/qemu/ooo/usr/local/bin/qemu-system-riscv64 \
-s -nographic -drive file=/home/song/8_riscv/3_acpi/Build_virt/RiscVVirtQemu/RELEASE_GCC5/FV/RISCV_VIRT.fd,if=pflash,format=raw,unit=1 \
-machine virt,acpi=off -smp 2 -m 2G \
-kernel /home/song/9_linux/linux/00_rv_def/arch/riscv/boot/Image \
-initrd /home/song/8_riscv/3_acpi/buildroot/output/images/rootfs.ext2 \
-append "root=/dev/ram ro console=ttyS0 earlycon=uart8250,mmio,0x10000000 efi=debug loglevel=8 memblock=debug" ## also panic by memtest
```
3. Then you will encounter the kernel panic logged in the above link

Other Information:

1. -------

This report is not identical to my prior report -- "kernel paniced when system hibernates" [1], but both of them
are closely related with the commit (3335068f8721 "riscv: Use PUD/P4D/PGD pages for the linear mapping").

With this commit, hibernation is trapped with "access fault" while accessing the PMP-protected regions (mmode_resv0@80000000)
from OpenSbi (BTW, hibernation is marked as nonportable by Conor[2]).

In this report, efi_init handoffs the memory mapping from Boot Services to memblock where reserves mmode_resv0@80000000,
so there is no "access fault" but "page fault".

And reverting commit 3335068f8721 indeed fixed this panic.

2. -------

As the gdb-pt-dump [3] tool shows, the PTE which covered the fault virtual address had the appropriate permission to store.
Is there another way to trigger the "Store/AMO page fault"? Or the creation of linear mapping in commit 3335068f8721 did something wrong?

```
(gdb) p/x $satp
$1 = 0xa000000000081708
(gdb) pt -satp 0xa000000000081708
Address : Length Permissions
0xff1bfffffea39000 : 0x1000 | W:1 X:0 R:1 S:1
0xff1bfffffebf9000 : 0x1000 | W:1 X:0 R:1 S:1
0xff1bfffffec00000 : 0x400000 | W:1 X:0 R:1 S:1
0xff60000000000000 : 0x1c0000 | W:1 X:0 R:1 S:1
0xff60000000200000 : 0xa00000 | W:0 X:0 R:1 S:1
0xff60000000c00000 : 0x7f000000 | W:1 X:0 R:1 S:1 // badaddr: ff6000007fdb1000
0xff6000007fdc0000 : 0x3d000 | W:1 X:0 R:1 S:1
0xff6000007ffbf000 : 0x1000 | W:1 X:0 R:1 S:1
0xffffffff80000000 : 0xc00000 | W:0 X:1 R:1 S:1
0xffffffff80c00000 : 0xa00000 | W:1 X:0 R:1 S:1

```

3. ------

You can also reproduce similar panic by appending "memtest" in kernel cmdline.
I have posted the memtest boot log at this link: https://termbin.com/1twl.

Please correct me if I'm wrong.

[1]: https://lore.kernel.org/linux-riscv/CAAYs2=gQvkhTeioMmqRDVGjdtNF_vhB+vm_1dHJxPNi75YDQ_Q@xxxxxxxxxxxxxx/
[2]: https://lore.kernel.org/linux-riscv/20230526-astride-detonator-9ae120051159@wendy/
[3]: https://github.com/martinradev/gdb-pt-dump

Thanks for the thorough report, really appreciated.

So there are multiple issues here:

- the first one is that the memory region for opensbi is marked as not
cacheable in the efi memory map, and then this region is not mapped in
the linear mapping:
[ 0.000000] efi: 0x000080000000-0x00008003ffff [Reserved | |
| | | | | | | | | | | |UC]

- the second one (that I feel a bit ashamed of...) is that I did not
check the alignment of the virtual address when choosing the map size
in best_map_size() and then we end up trying to map a physical region
aligned on 2MB that is actually not aligned on 2MB virtually because
the opensbi region is not mapped at all.

The issue 2 should be the root cause of this panic.

Here is my understanding of the necessity of the 2M-aligned VA for linear PMD mapping. Please correct me if I'm wrong.

I logged the `create_linear_mapping_range()` function.

```
song # lowmem region: [0x0000000081800000 -- 0x00000000ffe3d000], va: 0xff6000007fbc0000, pa: 0x00000000ffc00000, map_size: 200000 ,pg: e7
song # lowmem region: [0x0000000081800000 -- 0x00000000ffe3d000], va: 0xff6000007fdc0000, pa: 0x00000000ffe00000, map_size: 1000 ,pg: e7
```

The PA `0x00000000ffc00000` of this PMD mapping is aligned with PMD_SIZE but VA `0xff6000007fbc0000` is not.
After the `pmd_index()`, this 2M PA region is actually mapping the effective VA region `[0xff6000007fa00000,0xff6000007fc00000)`,
and any access of VA hole between the end of the effective VA region and the start VA of the next 4K mapping (`0xff6000007fdc0000`) will fault.

In this report, the memtest fault VA (`0xff6000007fc00000`) and the booting fault VA (`ff6000007fdb1000`) lie right in the VA hole.

When reverting the commit 3335068f8721, the kernel load address is always offseted by PMD_SIZE, kernel_map.va_pa_offset and
MIN_MEMBLOCK_ADDR follow it. So the linear PMD mapping will always take 2M-aligned VA. That's why this reverting works.

- the possible third one is that we should not map the linear mapping
using 4K pages, this would be slow in my opinion, and I think we
should waste a bit of memory to align va and pa on a 2MB boundary.
I also noticed this one.

So I'll fix the second issue, and possibly the third one, and if no
Thanks for your attention to this report, looking for your fixup.
one looks into why the opensbi region is mapped in UC, I'll take a
look at edk2.

Sorry for that,

Alex


--
Song Shuai
Thanks