Hi Shuai,
On 07/10/2023 08:28, Shuai Xue wrote:
Hardware errors could be signaled by synchronous interrupt,
I'm struggling with 'synchronous interrupt'. Do you mean arm64's 'precise' (all instructions before the exception were executed, and none after). Otherwise, surely any interrupt from a background scrubber is inherently asynchronous!
e.g. when an error is detected by a background scrubber, or signaled by synchronous exception, e.g. when an uncorrected error is consumed. Both synchronous and asynchronous error are queued and handled by a dedicated kthread in workqueue.
commit 7f17b4a121d0 ("ACPI: APEI: Kick the memory_failure() queue for synchronous errors") keep track of whether memory_failure() work was queued, and make task_work pending to flush out the workqueue so that the work for synchronous error is processed before returning to user-space.
It does it regardless, if user-space was interrupted by APEI any work queued as a result of that should be completed before we go back to user-space. Otherwise we can bounce between user-space and firmware, with the kernel only running the APEI code, and never making progress.
The trick ensures that the corrupted page is unmapped and poisoned. And after returning to user-space, the task starts at current instruction which triggering a page fault in which kernel will send SIGBUS to current process due to VM_FAULT_HWPOISON.
However, the memory failure recovery for hwpoison-aware mechanisms does not work as expected. For example, hwpoison-aware user-space processes like QEMU register their customized SIGBUS handler and enable early kill mode by seting PF_MCE_EARLY at initialization. Then the kernel will directy notify
(setting, directly)
the process by sending a SIGBUS signal in memory failure with wrong
si_code: the actual user-space process accessing the corrupt memory location, but its memory failure work is handled in a kthread context, so it will send SIGBUS with BUS_MCEERR_AO si_code to the actual user-space process instead of BUS_MCEERR_AR in kill_proc().
This is hard to parse, "the user-space process is accessing"? (dropping 'actual' and adding 'is')
Wasn't this behaviour fixed by the previous patch?
What problem are you fixing here?
To this end, separate synchronous and asynchronous error handling into different paths like X86 platform does:
- valid synchronous errors: queue a task_work to synchronously send SIGBUS before ret_to_user.
- valid asynchronous errors: queue a work into workqueue to asynchronously handle memory failure.
Why? The signal issue was fixed by the previous patch. Why delay the handling of a poisoned memory location further?
- abnormal branches such as invalid PA, unexpected severity, no memory failure config support, invalid GUID section, OOM, etc.
... do what?
Then for valid synchronous errors, the current context in memory failure is exactly belongs to the task consuming poison data and it will send SIBBUS with proper si_code.
diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c index 6f35f724cc14..1675ff77033d 100644 --- a/arch/x86/kernel/cpu/mce/core.c +++ b/arch/x86/kernel/cpu/mce/core.c @@ -1334,17 +1334,10 @@ static void kill_me_maybe(struct callback_head *cb) return; }
- /*
* -EHWPOISON from memory_failure() means that it already sent SIGBUS* to the current process with the proper error info,* -EOPNOTSUPP means hwpoison_filter() filtered the error event,** In both cases, no further processing is required.*/
- pr_err("Memory error not recovered");
- pr_err("Sending SIGBUS to current task due to memory error not recovered"); kill_me_now(cb);
}
I'm not sure how this hunk is relevant to the commit message.
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index 88178aa6222d..014401a65ed5 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -484,6 +497,18 @@ static bool ghes_do_memory_failure(u64 physical_addr, int flags) return false; }
- if (flags == MF_ACTION_REQUIRED && current->mm) {
twcb = kmalloc(sizeof(*twcb), GFP_ATOMIC);if (!twcb)return false;
Yuck - New failure modes! This is why the existing code always has this memory allocated in struct ghes_estatus_node.
twcb->pfn = pfn;twcb->flags = flags;init_task_work(&twcb->twork, memory_failure_cb);task_work_add(current, &twcb->twork, TWA_RESUME);return true;- }
- memory_failure_queue(pfn, flags); return true;
}
[..]
@@ -696,7 +721,14 @@ static bool ghes_do_proc(struct ghes *ghes, } }
- return queued;
- /*
* If no memory failure work is queued for abnormal synchronous* errors, do a force kill.*/- if (sync && !queued) {
pr_err("Sending SIGBUS to current task due to memory error not recovered");force_sig(SIGBUS);- }
}
I think this is a lot of churn, and this hunk is the the only meaningful change in behaviour. Can you explain how this happens?
Wouldn't it be simpler to split ghes_kick_task_work() to have a sync/async version. The synchronous version can unconditionally force_sig_mceerr(BUS_MCEERR_AR, ...) after memory_failure_queue_kick() - but that still means memory_failure() is unable to disappear errors that it fixed - see MF_RECOVERED.
diff --git a/mm/memory-failure.c b/mm/memory-failure.c index 4d6e43c88489..0d02f8a0b556 100644 --- a/mm/memory-failure.c +++ b/mm/memory-failure.c @@ -2161,9 +2161,12 @@ static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
- Must run in process context (e.g. a work queue) with interrupts
- enabled and no spinlocks held.
- Return: 0 for successfully handled the memory error,
-EOPNOTSUPP for hwpoison_filter() filtered the error event,
< 0(except -EOPNOTSUPP) on failure.
- Return values:
- 0 - success
- -EOPNOTSUPP - hwpoison_filter() filtered the error event.
- -EHWPOISON - sent SIGBUS to the current process with the proper
error info by kill_accessing_process().*/
- other negative values - failure
int memory_failure(unsigned long pfn, int flags) {
I'm not sure how this hunk is relevant to the commit message.
Thanks,
James