Re: [PATCH v13 3/3] ACPI: APEI: handle synchronous exceptions in task work

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri Sep 20, 2024 at 7:30 AM EEST, Shuai Xue wrote:
> The memory uncorrected error could be signaled by asynchronous interrupt
> (specifically, SPI in arm64 platform), e.g. when an error is detected by
> a background scrubber, or signaled by synchronous exception
> (specifically, data abort excepction in arm64 platform), e.g. when a CPU
> tries to access a poisoned cache line. Currently, both synchronous and
> asynchronous error use memory_failure_queue() to schedule
> memory_failure() exectute in kworker context.
>
> As a result, when a user-space process is accessing a poisoned data, a
> data abort is taken and the memory_failure() is executed in the kworker
> context:
>
>   - will send wrong si_code by SIGBUS signal in early_kill mode, and
>   - can not kill the user-space in some cases resulting a synchronous
>     error infinite loop
>
> Issue 1: send wrong si_code in early_kill mode
>
> Since commit a70297d22132 ("ACPI: APEI: set memory failure flags as
> MF_ACTION_REQUIRED on synchronous events")', the flag MF_ACTION_REQUIRED
> could be used to determine whether a synchronous exception occurs on
> ARM64 platform.  When a synchronous exception is detected, the kernel is
> expected to terminate the current process which has accessed poisoned
> page. This is done by sending a SIGBUS signal with an error code
> BUS_MCEERR_AR, indicating an action-required machine check error on
> read.
>
> However, when kill_proc() is called to terminate the processes who have
> the poisoned page mapped, it sends the incorrect SIGBUS error code
> BUS_MCEERR_AO because the context in which it operates is not the one
> where the error was triggered.
>
> To reproduce this problem:
>
>   #sysctl -w vm.memory_failure_early_kill=1
>   vm.memory_failure_early_kill = 1
>
>   # STEP2: inject an UCE error and consume it to trigger a synchronous error
>   #einj_mem_uc single
>   0: single   vaddr = 0xffffb0d75400 paddr = 4092d55b400
>   injecting ...
>   triggering ...
>   signal 7 code 5 addr 0xffffb0d75000
>   page not present
>   Test passed
>
> The si_code (code 5) from einj_mem_uc indicates that it is BUS_MCEERR_AO
> error and it is not fact.
>
> After this patch:
>
>   # STEP1: enable early kill mode
>   #sysctl -w vm.memory_failure_early_kill=1
>   vm.memory_failure_early_kill = 1
>   # STEP2: inject an UCE error and consume it to trigger a synchronous error
>   #einj_mem_uc single
>   0: single   vaddr = 0xffffb0d75400 paddr = 4092d55b400
>   injecting ...
>   triggering ...
>   signal 7 code 4 addr 0xffffb0d75000
>   page not present
>   Test passed
>
> The si_code (code 4) from einj_mem_uc indicates that it is BUS_MCEERR_AR
> error as we expected.
>
> Issue 2: a synchronous error infinite loop
>
> If a user-space process, e.g. devmem, a poisoned page which has been set
> HWPosion flag, kill_accessing_process() is called to send SIGBUS to the
> current processs with error info. Because the memory_failure() is
> executed in the kworker contex, it will just do nothing but return
> EFAULT. So, devmem will access the posioned page and trigger an
> excepction again, resulting in a synchronous error infinite loop. Such
> loop may cause platform firmware to exceed some threshold and reboot
> when Linux could have recovered from this error.
>
> To reproduce this problem:
>
>   # STEP 1: inject an UCE error, and kernel will set HWPosion flag for related page
>   #einj_mem_uc single
>   0: single   vaddr = 0xffffb0d75400 paddr = 4092d55b400
>   injecting ...
>   triggering ...
>   signal 7 code 4 addr 0xffffb0d75000
>   page not present
>   Test passed
>
>   # STEP 2: access the same page and it will trigger a synchronous error infinite loop
>   devmem 0x4092d55b400
>
> To fix above two issues, queue memory_failure() as a task_work so that it runs in
> the context of the process that is actually consuming the poisoned data.
>
> Signed-off-by: Shuai Xue <xueshuai@xxxxxxxxxxxxxxxxx>
> Tested-by: Ma Wupeng <mawupeng1@xxxxxxxxxx>
> Reviewed-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> Reviewed-by: Xiaofei Tan <tanxiaofei@xxxxxxxxxx>
> Reviewed-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> ---
>  drivers/acpi/apei/ghes.c | 78 +++++++++++++++++++++++-----------------
>  include/acpi/ghes.h      |  3 --
>  include/linux/mm.h       |  1 -
>  mm/memory-failure.c      | 13 -------
>  4 files changed, 45 insertions(+), 50 deletions(-)
>
> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
> index 93eb11482832..60d8044f14d1 100644
> --- a/drivers/acpi/apei/ghes.c
> +++ b/drivers/acpi/apei/ghes.c
> @@ -467,28 +467,42 @@ static void ghes_clear_estatus(struct ghes *ghes,
>  }
>  
>  /*
> - * Called as task_work before returning to user-space.
> - * Ensure any queued work has been done before we return to the context that
> - * triggered the notification.
> + * struct task_work - for synchronous RAS event
> + *
> + * @twork:                callback_head for task work
> + * @pfn:                  page frame number of corrupted page
> + * @flags:                work control flags
> + *
> + * Structure to pass task work to be handled before
> + * returning to user-space via task_work_add().
>   */
> -static void ghes_kick_task_work(struct callback_head *head)
> +struct task_work {
> +	struct callback_head twork;
> +	u64 pfn;
> +	int flags;
> +};

I'd rename this as ghes_task_work. It is too generic name IMHO, easily
confused with task_work.h definitions.

BR, Jarkko





[Index of Archives]     [Linux IBM ACPI]     [Linux Power Management]     [Linux Kernel]     [Linux Laptop]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]
  Powered by Linux