Re: BUG: drm/ast: soft lockup due to missing timeout in hardware polling (ast_2500_patch_ahb)

[Thomas Zimmermann <tzimmermann@suse.de>]

Hi

Am 13.05.26 um 09:36 schrieb w15303746062@163.com:
> From: Mingyu Wang <25181214217@stu.xidian.edu.cn>
>
> Hi all,
>
> While analyzing the AST driver's hardware interactions using our custom device emulation and fuzzing framework (DevGen), we observed a severe soft lockup in the DRM driver.
>
> Although the underlying trigger in our environment was an incomplete emulation of the ASPEED AHB bridge, this highlighted a critical defensive programming gap in the driver itself: a complete lack of timeout mechanisms in the hardware polling loops.
>
> This issue causes a complete system hang (CPU stuck for 143s+) and leads to subsequent I/O starvation and system paralysis (e.g., jbd2 and systemd-journald blocked).
>
> ### Crash Log Snippet
>
> watchdog: BUG: soft lockup - CPU#3 stuck for 143s! [systemd-udevd:162]
> RIP: 0010:ioread32+0x0/0xa0
> Call Trace:
>   <TASK>
>   __ast_read32
>   __ast_mindwm+0x4d/0x80 [ast]
>   ast_2500_patch_ahb+0x9d/0x120 [ast]
>   ast_detect_chip
>   ast_pci_probe+0xa1a/0xb70 [ast]
>   ...
>
> ### Vulnerability Analysis
>
> Even on bare-metal hardware, an unresponsive ASPEED chip, a PCIe bus fault, or unexpected hardware states can cause the kernel to hang forever in the following loops during initialization:
>
> 1. Inside `ast_2500_patch_ahb()`:
>      do {
>          __ast_moutdwm(regs, 0x1e6e2000, 0x1688A8A8);
>          data = __ast_mindwm(regs, 0x1e6e2000);
>      } while (data != 1); // <--- Infinite loop if hardware doesn't respond properly
>
> 2. Inside the underlying I/O accessors `__ast_mindwm()` and `__ast_moutdwm()`:
>      do {
>          data = __ast_read32(regs, 0xf004) & 0xffff0000;
>      } while (data != (r & 0xffff0000)); // <--- Infinite loop
>
> ### Proposed Fix Direction
>
> To prevent the kernel from hanging indefinitely and to gracefully abort the probe (`-ENODEV`) upon hardware failure, these loops must implement a timeout mechanism (e.g., using `readx_poll_timeout` or a loop counter with `udelay`).
>
> Interestingly, other functions in the same file (e.g., `mmc_test`) correctly implement a timeout counter (`if (++timeout > TIMEOUT) return false;`), but the initialization paths mentioned above blindly trust the hardware state.
>
> We are reporting this defect so that the maintainers can decide the appropriate timeout thresholds and implement a safe fallback mechanism across the `ast` driver.
>
> Please let us know if you need more information or the full dmesg log.

Thanks for reporting. Will you send a patch?

Best regards
Thomas

>
> Reported-by: Mingyu Wang <25181214217@stu.xidian.edu.cn>
>
> Best regards,
> Mingyu Wang
>

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)