Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Jonathan Cameron]

On Wed, 19 Jun 2024 00:53:10 +0800
Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:

> Background:
> Since CXL device is a memory device, while CPU consumes a poison page of 
> CXL device, it always triggers a MCE by interrupt (INT18), no matter 
> which-First path is configured.  This is the first report.  Then 
> currently, in FW-First path, the poison event is transferred according 
> to the following process: CXL device -> firmware -> OS:ACPI->APEI->GHES 
>  -> CPER -> trace report.  This is the second one.  These two reports  
> are indicating the same poisoning page, which is the so-called "duplicate
> report"[1].  And the memory_failure() handling I'm trying to add in
> OS-First path could also be another duplicate report.
> 
> Hope the flow below could make it easier to understand:
> CPU accesses bad memory on CXL device, then
>  -> MCE (INT18), *always* report (1)
>  -> * FW-First (implemented now)
>       -> CXL device -> FW
> 	      -> OS:ACPI->APEI->GHES->CPER -> trace report (2.a)  
>     * OS-First (not implemented yet, I'm working on it)
>       -> CXL device -> MSI
> 	      -> OS:CXL driver -> memory_failure() (2.b)  
> so, the (1) and (2.a/b) are duplicated.
> 
> (I didn't get response in my reply for [1] while I have to make patch to
> solve this problem, so please correct me if my understanding is wrong.)
> 
> This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
> to check whether the current poison page has been reported (if yes,
> stop the notifier chain, won't call the following memory_failure()
> to report), into `x86_mce_decoder_chain`.  In this way, if the poison
> page already handled(recorded and reported) in (1) or (2), the other one
> won't duplicate the report.  The record could be clear when
> cxl_clear_poison() is called.
> 
> [1] https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/
> 
> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>

So poison can be cleared in a number of ways and a CXL poison clear command
is unfortunately only one of them.  Some architectures have instructions
that guarantee to write a whole cacheline and can clear things as well.
I believe x86 does for starters.

+CC linux-edac and related maintainers / reviewers.
    linux-mm and hwpoison maintainer.

So I think this needs a more general solution that encompasses 
more general cleanup of poison.

Trivial comments inline.

Jonathan


> ---
>  arch/x86/include/asm/mce.h |   1 +
>  drivers/cxl/core/mbox.c    | 130 +++++++++++++++++++++++++++++++++++++
>  drivers/cxl/core/memdev.c  |   6 +-
>  drivers/cxl/cxlmem.h       |   3 +
>  4 files changed, 139 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
> index dfd2e9699bd7..d8109c48e7d9 100644
> --- a/arch/x86/include/asm/mce.h
> +++ b/arch/x86/include/asm/mce.h
> @@ -182,6 +182,7 @@ enum mce_notifier_prios {
>  	MCE_PRIO_NFIT,
>  	MCE_PRIO_EXTLOG,
>  	MCE_PRIO_UC,
> +	MCE_PRIO_CXL,
>  	MCE_PRIO_EARLY,
>  	MCE_PRIO_CEC,
>  	MCE_PRIO_HIGHEST = MCE_PRIO_CEC
> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> index 2626f3fff201..0eb3c5401e81 100644
> --- a/drivers/cxl/core/mbox.c
> +++ b/drivers/cxl/core/mbox.c
> @@ -4,6 +4,8 @@
>  #include <linux/debugfs.h>
>  #include <linux/ktime.h>
>  #include <linux/mutex.h>
> +#include <linux/notifier.h>
> +#include <asm/mce.h>
>  #include <asm/unaligned.h>
>  #include <cxlpci.h>
>  #include <cxlmem.h>
> @@ -880,6 +882,9 @@ void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>  		if (cxlr)
>  			hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>  
> +		if (hpa != ULLONG_MAX && cxl_mce_recorded(hpa))
> +			return;
> +
>  		if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
>  			trace_cxl_general_media(cxlmd, type, cxlr, hpa,
>  						&evt->gen_media);
> @@ -1408,6 +1413,127 @@ int cxl_poison_state_init(struct cxl_memdev_state *mds)
>  }
>  EXPORT_SYMBOL_NS_GPL(cxl_poison_state_init, CXL);
>  
> +struct cxl_mce_record {
> +	struct list_head node;
> +	u64 hpa;
> +};
> +LIST_HEAD(cxl_mce_records);
> +DEFINE_MUTEX(cxl_mce_mutex);
> +
> +bool cxl_mce_recorded(u64 hpa)
> +{
> +	struct cxl_mce_record *cur, *next, *rec;
> +	int rc;
> +
> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);

guard(mutex)(&cxl_mce_muted);

> +	if (rc)
> +		return false;
> +
> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
> +		if (cur->hpa == hpa) {
> +			mutex_unlock(&cxl_mce_mutex);
> +			return true;
> +		}
> +	}
> +
> +	rec = kmalloc(sizeof(struct cxl_mce_record), GFP_KERNEL);
> +	rec->hpa = hpa;
> +	list_add(&cxl_mce_records, &rec->node);
> +
> +	mutex_unlock(&cxl_mce_mutex);
> +
> +	return false;
> +}
> +
> +void cxl_mce_clear(u64 hpa)
> +{
> +	struct cxl_mce_record *cur, *next;
> +	int rc;
> +
> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);

Maybe cond_guard().

> +	if (rc)
> +		return;
> +
> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
> +		if (cur->hpa == hpa) {
> +			list_del(&cur->node);
> +			break;
> +		}
> +	}
> +
> +	mutex_unlock(&cxl_mce_mutex);
> +}
> +
> +struct cxl_contains_hpa_context {
> +	bool contains;
> +	u64 hpa;
> +};
> +
> +static int __cxl_contains_hpa(struct device *dev, void *arg)
> +{
> +	struct cxl_contains_hpa_context *ctx = arg;
> +	struct cxl_endpoint_decoder *cxled;
> +	struct range *range;
> +	u64 hpa = ctx->hpa;
> +
> +	if (!is_endpoint_decoder(dev))
> +		return 0;
> +
> +	cxled = to_cxl_endpoint_decoder(dev);
> +	range = &cxled->cxld.hpa_range;
> +
> +	if (range->start <= hpa && hpa <= range->end) {
> +		ctx->contains = true;
> +		return 1;
> +	}
> +
> +	return 0;
> +}
> +
> +static bool cxl_contains_hpa(const struct cxl_memdev *cxlmd, u64 hpa)
> +{
> +	struct cxl_contains_hpa_context ctx = {
> +		.contains = false,
> +		.hpa = hpa,
> +	};
> +	struct cxl_port *port;
> +
> +	port = cxlmd->endpoint;
> +	if (port && is_cxl_endpoint(port) && cxl_num_decoders_committed(port))
> +		device_for_each_child(&port->dev, &ctx, __cxl_contains_hpa);
> +
> +	return ctx.contains;
> +}
> +
> +static int cxl_handle_mce(struct notifier_block *nb, unsigned long val,
> +			  void *data)
> +{
> +	struct mce *mce = (struct mce *)data;
> +	struct cxl_memdev_state *mds = container_of(nb, struct cxl_memdev_state,
> +						    mce_notifier);
> +	u64 hpa;
> +
> +	if (!mce || !mce_usable_address(mce))
> +		return NOTIFY_DONE;
> +
> +	hpa = mce->addr & MCI_ADDR_PHYSADDR;
> +
> +	/* Check if the PFN is located on this CXL device */
> +	if (!pfn_valid(hpa >> PAGE_SHIFT) &&
> +	    !cxl_contains_hpa(mds->cxlds.cxlmd, hpa))
> +		return NOTIFY_DONE;
> +
> +	/*
> +	 * Search PFN in the cxl_mce_records, if already exists, don't continue
> +	 * to do memory_failure() to avoid a poison address being reported
> +	 * more than once.
> +	 */
> +	if (cxl_mce_recorded(hpa))
> +		return NOTIFY_STOP;
> +	else
> +		return NOTIFY_OK;
> +}
> +
>  struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
>  {
>  	struct cxl_memdev_state *mds;
> @@ -1427,6 +1553,10 @@ struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
>  	mds->ram_perf.qos_class = CXL_QOS_CLASS_INVALID;
>  	mds->pmem_perf.qos_class = CXL_QOS_CLASS_INVALID;
>  
> +	mds->mce_notifier.notifier_call = cxl_handle_mce;
> +	mds->mce_notifier.priority = MCE_PRIO_CXL;
> +	mce_register_decode_chain(&mds->mce_notifier);
> +
>  	return mds;
>  }
>  EXPORT_SYMBOL_NS_GPL(cxl_memdev_state_create, CXL);
> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> index 0277726afd04..aa3ac89d17be 100644
> --- a/drivers/cxl/core/memdev.c
> +++ b/drivers/cxl/core/memdev.c
> @@ -376,10 +376,14 @@ int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa)
>  		goto out;
>  
>  	cxlr = cxl_dpa_to_region(cxlmd, dpa);
> -	if (cxlr)
> +	if (cxlr) {
> +		u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> +
> +		cxl_mce_clear(hpa);
>  		dev_warn_once(mds->cxlds.dev,
>  			      "poison clear dpa:%#llx region: %s\n", dpa,
>  			      dev_name(&cxlr->dev));
> +	}
>  
>  	record = (struct cxl_poison_record) {
>  		.address = cpu_to_le64(dpa),
> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
> index 19aba81cdf13..fbf8d9f46984 100644
> --- a/drivers/cxl/cxlmem.h
> +++ b/drivers/cxl/cxlmem.h
> @@ -501,6 +501,7 @@ struct cxl_memdev_state {
>  	struct cxl_fw_state fw;
>  
>  	struct rcuwait mbox_wait;
> +	struct notifier_block mce_notifier;
>  	int (*mbox_send)(struct cxl_memdev_state *mds,
>  			 struct cxl_mbox_cmd *cmd);
>  };
> @@ -836,6 +837,8 @@ int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
>  int cxl_trigger_poison_list(struct cxl_memdev *cxlmd);
>  int cxl_inject_poison(struct cxl_memdev *cxlmd, u64 dpa);
>  int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa);
> +bool cxl_mce_recorded(u64 pfn);
> +void cxl_mce_clear(u64 pfn);
>  
>  #ifdef CONFIG_CXL_SUSPEND
>  void cxl_mem_active_inc(void);

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Shiyang Ruan]

在 2024/6/21 1:02, Jonathan Cameron 写道:
> On Wed, 19 Jun 2024 00:53:10 +0800
> Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:
> 
>> Background:
>> Since CXL device is a memory device, while CPU consumes a poison page of
>> CXL device, it always triggers a MCE by interrupt (INT18), no matter
>> which-First path is configured.  This is the first report.  Then
>> currently, in FW-First path, the poison event is transferred according
>> to the following process: CXL device -> firmware -> OS:ACPI->APEI->GHES
>>   -> CPER -> trace report.  This is the second one.  These two reports
>> are indicating the same poisoning page, which is the so-called "duplicate
>> report"[1].  And the memory_failure() handling I'm trying to add in
>> OS-First path could also be another duplicate report.
>>
>> Hope the flow below could make it easier to understand:
>> CPU accesses bad memory on CXL device, then
>>   -> MCE (INT18), *always* report (1)
>>   -> * FW-First (implemented now)
>>        -> CXL device -> FW
>> 	      -> OS:ACPI->APEI->GHES->CPER -> trace report (2.a)
>>      * OS-First (not implemented yet, I'm working on it)
>>        -> CXL device -> MSI
>> 	      -> OS:CXL driver -> memory_failure() (2.b)
>> so, the (1) and (2.a/b) are duplicated.
>>
>> (I didn't get response in my reply for [1] while I have to make patch to
>> solve this problem, so please correct me if my understanding is wrong.)
>>
>> This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
>> to check whether the current poison page has been reported (if yes,
>> stop the notifier chain, won't call the following memory_failure()
>> to report), into `x86_mce_decoder_chain`.  In this way, if the poison
>> page already handled(recorded and reported) in (1) or (2), the other one
>> won't duplicate the report.  The record could be clear when
>> cxl_clear_poison() is called.
>>
>> [1] https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/
>>
>> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> 
> So poison can be cleared in a number of ways and a CXL poison clear command
> is unfortunately only one of them.  Some architectures have instructions
> that guarantee to write a whole cacheline and can clear things as well.
> I believe x86 does for starters.

According to the CXL Spec, to clear an error record on device, an 
explicit clear operation is required (I think this means sending a mbox 
command).  I'm not sure if it is able to clear device error by just 
writing a whole cacheline.

> 
> +CC linux-edac and related maintainers / reviewers.
>      linux-mm and hwpoison maintainer.
> 
> So I think this needs a more general solution that encompasses
> more general cleanup of poison.
> 
> Trivial comments inline.

Thanks

> 
> Jonathan
> 
> 
>> ---
>>   arch/x86/include/asm/mce.h |   1 +
>>   drivers/cxl/core/mbox.c    | 130 +++++++++++++++++++++++++++++++++++++
>>   drivers/cxl/core/memdev.c  |   6 +-
>>   drivers/cxl/cxlmem.h       |   3 +
>>   4 files changed, 139 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
>> index dfd2e9699bd7..d8109c48e7d9 100644
>> --- a/arch/x86/include/asm/mce.h
>> +++ b/arch/x86/include/asm/mce.h
>> @@ -182,6 +182,7 @@ enum mce_notifier_prios {
>>   	MCE_PRIO_NFIT,
>>   	MCE_PRIO_EXTLOG,
>>   	MCE_PRIO_UC,
>> +	MCE_PRIO_CXL,
>>   	MCE_PRIO_EARLY,
>>   	MCE_PRIO_CEC,
>>   	MCE_PRIO_HIGHEST = MCE_PRIO_CEC
>> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
>> index 2626f3fff201..0eb3c5401e81 100644
>> --- a/drivers/cxl/core/mbox.c
>> +++ b/drivers/cxl/core/mbox.c
>> @@ -4,6 +4,8 @@
>>   #include <linux/debugfs.h>
>>   #include <linux/ktime.h>
>>   #include <linux/mutex.h>
>> +#include <linux/notifier.h>
>> +#include <asm/mce.h>
>>   #include <asm/unaligned.h>
>>   #include <cxlpci.h>
>>   #include <cxlmem.h>
>> @@ -880,6 +882,9 @@ void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>>   		if (cxlr)
>>   			hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>>   
>> +		if (hpa != ULLONG_MAX && cxl_mce_recorded(hpa))
>> +			return;
>> +
>>   		if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
>>   			trace_cxl_general_media(cxlmd, type, cxlr, hpa,
>>   						&evt->gen_media);
>> @@ -1408,6 +1413,127 @@ int cxl_poison_state_init(struct cxl_memdev_state *mds)
>>   }
>>   EXPORT_SYMBOL_NS_GPL(cxl_poison_state_init, CXL);
>>   
>> +struct cxl_mce_record {
>> +	struct list_head node;
>> +	u64 hpa;
>> +};
>> +LIST_HEAD(cxl_mce_records);
>> +DEFINE_MUTEX(cxl_mce_mutex);
>> +
>> +bool cxl_mce_recorded(u64 hpa)
>> +{
>> +	struct cxl_mce_record *cur, *next, *rec;
>> +	int rc;
>> +
>> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);
> 
> guard(mutex)(&cxl_mce_muted);
> 
>> +	if (rc)
>> +		return false;
>> +
>> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
>> +		if (cur->hpa == hpa) {
>> +			mutex_unlock(&cxl_mce_mutex);
>> +			return true;
>> +		}
>> +	}
>> +
>> +	rec = kmalloc(sizeof(struct cxl_mce_record), GFP_KERNEL);
>> +	rec->hpa = hpa;
>> +	list_add(&cxl_mce_records, &rec->node);
>> +
>> +	mutex_unlock(&cxl_mce_mutex);
>> +
>> +	return false;
>> +}
>> +
>> +void cxl_mce_clear(u64 hpa)
>> +{
>> +	struct cxl_mce_record *cur, *next;
>> +	int rc;
>> +
>> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);
> 
> Maybe cond_guard().

Ok, this is better.  I'll use automatic clean locks instead.


--
Thanks,
Ruan.

> 
>> +	if (rc)
>> +		return;
>> +
>> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
>> +		if (cur->hpa == hpa) {
>> +			list_del(&cur->node);
>> +			break;
>> +		}
>> +	}
>> +
>> +	mutex_unlock(&cxl_mce_mutex);
>> +}
>> +
>> +struct cxl_contains_hpa_context {
>> +	bool contains;
>> +	u64 hpa;
>> +};
>> +
>> +static int __cxl_contains_hpa(struct device *dev, void *arg)
>> +{
>> +	struct cxl_contains_hpa_context *ctx = arg;
>> +	struct cxl_endpoint_decoder *cxled;
>> +	struct range *range;
>> +	u64 hpa = ctx->hpa;
>> +
>> +	if (!is_endpoint_decoder(dev))
>> +		return 0;
>> +
>> +	cxled = to_cxl_endpoint_decoder(dev);
>> +	range = &cxled->cxld.hpa_range;
>> +
>> +	if (range->start <= hpa && hpa <= range->end) {
>> +		ctx->contains = true;
>> +		return 1;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +static bool cxl_contains_hpa(const struct cxl_memdev *cxlmd, u64 hpa)
>> +{
>> +	struct cxl_contains_hpa_context ctx = {
>> +		.contains = false,
>> +		.hpa = hpa,
>> +	};
>> +	struct cxl_port *port;
>> +
>> +	port = cxlmd->endpoint;
>> +	if (port && is_cxl_endpoint(port) && cxl_num_decoders_committed(port))
>> +		device_for_each_child(&port->dev, &ctx, __cxl_contains_hpa);
>> +
>> +	return ctx.contains;
>> +}
>> +
>> +static int cxl_handle_mce(struct notifier_block *nb, unsigned long val,
>> +			  void *data)
>> +{
>> +	struct mce *mce = (struct mce *)data;
>> +	struct cxl_memdev_state *mds = container_of(nb, struct cxl_memdev_state,
>> +						    mce_notifier);
>> +	u64 hpa;
>> +
>> +	if (!mce || !mce_usable_address(mce))
>> +		return NOTIFY_DONE;
>> +
>> +	hpa = mce->addr & MCI_ADDR_PHYSADDR;
>> +
>> +	/* Check if the PFN is located on this CXL device */
>> +	if (!pfn_valid(hpa >> PAGE_SHIFT) &&
>> +	    !cxl_contains_hpa(mds->cxlds.cxlmd, hpa))
>> +		return NOTIFY_DONE;
>> +
>> +	/*
>> +	 * Search PFN in the cxl_mce_records, if already exists, don't continue
>> +	 * to do memory_failure() to avoid a poison address being reported
>> +	 * more than once.
>> +	 */
>> +	if (cxl_mce_recorded(hpa))
>> +		return NOTIFY_STOP;
>> +	else
>> +		return NOTIFY_OK;
>> +}
>> +
>>   struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
>>   {
>>   	struct cxl_memdev_state *mds;
>> @@ -1427,6 +1553,10 @@ struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
>>   	mds->ram_perf.qos_class = CXL_QOS_CLASS_INVALID;
>>   	mds->pmem_perf.qos_class = CXL_QOS_CLASS_INVALID;
>>   
>> +	mds->mce_notifier.notifier_call = cxl_handle_mce;
>> +	mds->mce_notifier.priority = MCE_PRIO_CXL;
>> +	mce_register_decode_chain(&mds->mce_notifier);
>> +
>>   	return mds;
>>   }
>>   EXPORT_SYMBOL_NS_GPL(cxl_memdev_state_create, CXL);
>> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
>> index 0277726afd04..aa3ac89d17be 100644
>> --- a/drivers/cxl/core/memdev.c
>> +++ b/drivers/cxl/core/memdev.c
>> @@ -376,10 +376,14 @@ int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa)
>>   		goto out;
>>   
>>   	cxlr = cxl_dpa_to_region(cxlmd, dpa);
>> -	if (cxlr)
>> +	if (cxlr) {
>> +		u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>> +
>> +		cxl_mce_clear(hpa);
>>   		dev_warn_once(mds->cxlds.dev,
>>   			      "poison clear dpa:%#llx region: %s\n", dpa,
>>   			      dev_name(&cxlr->dev));
>> +	}
>>   
>>   	record = (struct cxl_poison_record) {
>>   		.address = cpu_to_le64(dpa),
>> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
>> index 19aba81cdf13..fbf8d9f46984 100644
>> --- a/drivers/cxl/cxlmem.h
>> +++ b/drivers/cxl/cxlmem.h
>> @@ -501,6 +501,7 @@ struct cxl_memdev_state {
>>   	struct cxl_fw_state fw;
>>   
>>   	struct rcuwait mbox_wait;
>> +	struct notifier_block mce_notifier;
>>   	int (*mbox_send)(struct cxl_memdev_state *mds,
>>   			 struct cxl_mbox_cmd *cmd);
>>   };
>> @@ -836,6 +837,8 @@ int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
>>   int cxl_trigger_poison_list(struct cxl_memdev *cxlmd);
>>   int cxl_inject_poison(struct cxl_memdev *cxlmd, u64 dpa);
>>   int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa);
>> +bool cxl_mce_recorded(u64 pfn);
>> +void cxl_mce_clear(u64 pfn);
>>   
>>   #ifdef CONFIG_CXL_SUSPEND
>>   void cxl_mem_active_inc(void);
> 

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Jonathan Cameron]

On Fri, 21 Jun 2024 18:16:33 +0800
Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:

> 在 2024/6/21 1:02, Jonathan Cameron 写道:
> > On Wed, 19 Jun 2024 00:53:10 +0800
> > Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:
> >   
> >> Background:
> >> Since CXL device is a memory device, while CPU consumes a poison page of
> >> CXL device, it always triggers a MCE by interrupt (INT18), no matter
> >> which-First path is configured.  This is the first report.  Then
> >> currently, in FW-First path, the poison event is transferred according
> >> to the following process: CXL device -> firmware -> OS:ACPI->APEI->GHES  
> >>   -> CPER -> trace report.  This is the second one.  These two reports  
> >> are indicating the same poisoning page, which is the so-called "duplicate
> >> report"[1].  And the memory_failure() handling I'm trying to add in
> >> OS-First path could also be another duplicate report.
> >>
> >> Hope the flow below could make it easier to understand:
> >> CPU accesses bad memory on CXL device, then  
> >>   -> MCE (INT18), *always* report (1)
> >>   -> * FW-First (implemented now)
> >>        -> CXL device -> FW
> >> 	      -> OS:ACPI->APEI->GHES->CPER -> trace report (2.a)  
> >>      * OS-First (not implemented yet, I'm working on it)  
> >>        -> CXL device -> MSI
> >> 	      -> OS:CXL driver -> memory_failure() (2.b)  
> >> so, the (1) and (2.a/b) are duplicated.
> >>
> >> (I didn't get response in my reply for [1] while I have to make patch to
> >> solve this problem, so please correct me if my understanding is wrong.)
> >>
> >> This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
> >> to check whether the current poison page has been reported (if yes,
> >> stop the notifier chain, won't call the following memory_failure()
> >> to report), into `x86_mce_decoder_chain`.  In this way, if the poison
> >> page already handled(recorded and reported) in (1) or (2), the other one
> >> won't duplicate the report.  The record could be clear when
> >> cxl_clear_poison() is called.
> >>
> >> [1] https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/
> >>
> >> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>  
> > 
> > So poison can be cleared in a number of ways and a CXL poison clear command
> > is unfortunately only one of them.  Some architectures have instructions
> > that guarantee to write a whole cacheline and can clear things as well.
> > I believe x86 does for starters.  
> 
> According to the CXL Spec, to clear an error record on device, an 
> explicit clear operation is required (I think this means sending a mbox 
> command).  I'm not sure if it is able to clear device error by just 
> writing a whole cacheline.
> 

Please give a spec reference.  The only one I'm immediately seeing is
in 8.3.9.9.4.1 Get Poison List (opcode 43000h)
which says
"When poison is cleared"
but doesn't talk about how.

For TSP cases Clear poison is not allowed, so if they want to clear it
they will have to do it a suitable CPU arch approach not that command
(which may not be implemented in a given device - I gather it is
 awkward to do and a backdoor from control path to datapath isn't
 a popular feature!).

+CC John Groves.  John, any info you can share on whether you expect all
devices with a poison list to support the clear poison command?



> > 
> > +CC linux-edac and related maintainers / reviewers.
> >      linux-mm and hwpoison maintainer.
> > 
> > So I think this needs a more general solution that encompasses
> > more general cleanup of poison.
> > 
> > Trivial comments inline.  
> 
> Thanks
> 
> > 
> > Jonathan
> > 
> >   
> >> ---
> >>   arch/x86/include/asm/mce.h |   1 +
> >>   drivers/cxl/core/mbox.c    | 130 +++++++++++++++++++++++++++++++++++++
> >>   drivers/cxl/core/memdev.c  |   6 +-
> >>   drivers/cxl/cxlmem.h       |   3 +
> >>   4 files changed, 139 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
> >> index dfd2e9699bd7..d8109c48e7d9 100644
> >> --- a/arch/x86/include/asm/mce.h
> >> +++ b/arch/x86/include/asm/mce.h
> >> @@ -182,6 +182,7 @@ enum mce_notifier_prios {
> >>   	MCE_PRIO_NFIT,
> >>   	MCE_PRIO_EXTLOG,
> >>   	MCE_PRIO_UC,
> >> +	MCE_PRIO_CXL,
> >>   	MCE_PRIO_EARLY,
> >>   	MCE_PRIO_CEC,
> >>   	MCE_PRIO_HIGHEST = MCE_PRIO_CEC
> >> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> >> index 2626f3fff201..0eb3c5401e81 100644
> >> --- a/drivers/cxl/core/mbox.c
> >> +++ b/drivers/cxl/core/mbox.c
> >> @@ -4,6 +4,8 @@
> >>   #include <linux/debugfs.h>
> >>   #include <linux/ktime.h>
> >>   #include <linux/mutex.h>
> >> +#include <linux/notifier.h>
> >> +#include <asm/mce.h>
> >>   #include <asm/unaligned.h>
> >>   #include <cxlpci.h>
> >>   #include <cxlmem.h>
> >> @@ -880,6 +882,9 @@ void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> >>   		if (cxlr)
> >>   			hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> >>   
> >> +		if (hpa != ULLONG_MAX && cxl_mce_recorded(hpa))
> >> +			return;
> >> +
> >>   		if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> >>   			trace_cxl_general_media(cxlmd, type, cxlr, hpa,
> >>   						&evt->gen_media);
> >> @@ -1408,6 +1413,127 @@ int cxl_poison_state_init(struct cxl_memdev_state *mds)
> >>   }
> >>   EXPORT_SYMBOL_NS_GPL(cxl_poison_state_init, CXL);
> >>   
> >> +struct cxl_mce_record {
> >> +	struct list_head node;
> >> +	u64 hpa;
> >> +};
> >> +LIST_HEAD(cxl_mce_records);
> >> +DEFINE_MUTEX(cxl_mce_mutex);
> >> +
> >> +bool cxl_mce_recorded(u64 hpa)
> >> +{
> >> +	struct cxl_mce_record *cur, *next, *rec;
> >> +	int rc;
> >> +
> >> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);  
> > 
> > guard(mutex)(&cxl_mce_muted);
> >   
> >> +	if (rc)
> >> +		return false;
> >> +
> >> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
> >> +		if (cur->hpa == hpa) {
> >> +			mutex_unlock(&cxl_mce_mutex);
> >> +			return true;
> >> +		}
> >> +	}
> >> +
> >> +	rec = kmalloc(sizeof(struct cxl_mce_record), GFP_KERNEL);
> >> +	rec->hpa = hpa;
> >> +	list_add(&cxl_mce_records, &rec->node);
> >> +
> >> +	mutex_unlock(&cxl_mce_mutex);
> >> +
> >> +	return false;
> >> +}
> >> +
> >> +void cxl_mce_clear(u64 hpa)
> >> +{
> >> +	struct cxl_mce_record *cur, *next;
> >> +	int rc;
> >> +
> >> +	rc = mutex_lock_interruptible(&cxl_mce_mutex);  
> > 
> > Maybe cond_guard().  
> 
> Ok, this is better.  I'll use automatic clean locks instead.
> 
> 
> --
> Thanks,
> Ruan.
> 
> >   
> >> +	if (rc)
> >> +		return;
> >> +
> >> +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
> >> +		if (cur->hpa == hpa) {
> >> +			list_del(&cur->node);
> >> +			break;
> >> +		}
> >> +	}
> >> +
> >> +	mutex_unlock(&cxl_mce_mutex);
> >> +}
> >> +
> >> +struct cxl_contains_hpa_context {
> >> +	bool contains;
> >> +	u64 hpa;
> >> +};
> >> +
> >> +static int __cxl_contains_hpa(struct device *dev, void *arg)
> >> +{
> >> +	struct cxl_contains_hpa_context *ctx = arg;
> >> +	struct cxl_endpoint_decoder *cxled;
> >> +	struct range *range;
> >> +	u64 hpa = ctx->hpa;
> >> +
> >> +	if (!is_endpoint_decoder(dev))
> >> +		return 0;
> >> +
> >> +	cxled = to_cxl_endpoint_decoder(dev);
> >> +	range = &cxled->cxld.hpa_range;
> >> +
> >> +	if (range->start <= hpa && hpa <= range->end) {
> >> +		ctx->contains = true;
> >> +		return 1;
> >> +	}
> >> +
> >> +	return 0;
> >> +}
> >> +
> >> +static bool cxl_contains_hpa(const struct cxl_memdev *cxlmd, u64 hpa)
> >> +{
> >> +	struct cxl_contains_hpa_context ctx = {
> >> +		.contains = false,
> >> +		.hpa = hpa,
> >> +	};
> >> +	struct cxl_port *port;
> >> +
> >> +	port = cxlmd->endpoint;
> >> +	if (port && is_cxl_endpoint(port) && cxl_num_decoders_committed(port))
> >> +		device_for_each_child(&port->dev, &ctx, __cxl_contains_hpa);
> >> +
> >> +	return ctx.contains;
> >> +}
> >> +
> >> +static int cxl_handle_mce(struct notifier_block *nb, unsigned long val,
> >> +			  void *data)
> >> +{
> >> +	struct mce *mce = (struct mce *)data;
> >> +	struct cxl_memdev_state *mds = container_of(nb, struct cxl_memdev_state,
> >> +						    mce_notifier);
> >> +	u64 hpa;
> >> +
> >> +	if (!mce || !mce_usable_address(mce))
> >> +		return NOTIFY_DONE;
> >> +
> >> +	hpa = mce->addr & MCI_ADDR_PHYSADDR;
> >> +
> >> +	/* Check if the PFN is located on this CXL device */
> >> +	if (!pfn_valid(hpa >> PAGE_SHIFT) &&
> >> +	    !cxl_contains_hpa(mds->cxlds.cxlmd, hpa))
> >> +		return NOTIFY_DONE;
> >> +
> >> +	/*
> >> +	 * Search PFN in the cxl_mce_records, if already exists, don't continue
> >> +	 * to do memory_failure() to avoid a poison address being reported
> >> +	 * more than once.
> >> +	 */
> >> +	if (cxl_mce_recorded(hpa))
> >> +		return NOTIFY_STOP;
> >> +	else
> >> +		return NOTIFY_OK;
> >> +}
> >> +
> >>   struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
> >>   {
> >>   	struct cxl_memdev_state *mds;
> >> @@ -1427,6 +1553,10 @@ struct cxl_memdev_state *cxl_memdev_state_create(struct device *dev)
> >>   	mds->ram_perf.qos_class = CXL_QOS_CLASS_INVALID;
> >>   	mds->pmem_perf.qos_class = CXL_QOS_CLASS_INVALID;
> >>   
> >> +	mds->mce_notifier.notifier_call = cxl_handle_mce;
> >> +	mds->mce_notifier.priority = MCE_PRIO_CXL;
> >> +	mce_register_decode_chain(&mds->mce_notifier);
> >> +
> >>   	return mds;
> >>   }
> >>   EXPORT_SYMBOL_NS_GPL(cxl_memdev_state_create, CXL);
> >> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> >> index 0277726afd04..aa3ac89d17be 100644
> >> --- a/drivers/cxl/core/memdev.c
> >> +++ b/drivers/cxl/core/memdev.c
> >> @@ -376,10 +376,14 @@ int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa)
> >>   		goto out;
> >>   
> >>   	cxlr = cxl_dpa_to_region(cxlmd, dpa);
> >> -	if (cxlr)
> >> +	if (cxlr) {
> >> +		u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> >> +
> >> +		cxl_mce_clear(hpa);
> >>   		dev_warn_once(mds->cxlds.dev,
> >>   			      "poison clear dpa:%#llx region: %s\n", dpa,
> >>   			      dev_name(&cxlr->dev));
> >> +	}
> >>   
> >>   	record = (struct cxl_poison_record) {
> >>   		.address = cpu_to_le64(dpa),
> >> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
> >> index 19aba81cdf13..fbf8d9f46984 100644
> >> --- a/drivers/cxl/cxlmem.h
> >> +++ b/drivers/cxl/cxlmem.h
> >> @@ -501,6 +501,7 @@ struct cxl_memdev_state {
> >>   	struct cxl_fw_state fw;
> >>   
> >>   	struct rcuwait mbox_wait;
> >> +	struct notifier_block mce_notifier;
> >>   	int (*mbox_send)(struct cxl_memdev_state *mds,
> >>   			 struct cxl_mbox_cmd *cmd);
> >>   };
> >> @@ -836,6 +837,8 @@ int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
> >>   int cxl_trigger_poison_list(struct cxl_memdev *cxlmd);
> >>   int cxl_inject_poison(struct cxl_memdev *cxlmd, u64 dpa);
> >>   int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa);
> >> +bool cxl_mce_recorded(u64 pfn);
> >> +void cxl_mce_clear(u64 pfn);
> >>   
> >>   #ifdef CONFIG_CXL_SUSPEND
> >>   void cxl_mem_active_inc(void);  
> >   

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Dan Williams]

Jonathan Cameron wrote:
> On Wed, 19 Jun 2024 00:53:10 +0800
> Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:
> 
> > Background:
> > Since CXL device is a memory device, while CPU consumes a poison page of 
> > CXL device, it always triggers a MCE by interrupt (INT18), no matter 
> > which-First path is configured.  This is the first report.  Then 
> > currently, in FW-First path, the poison event is transferred according 
> > to the following process: CXL device -> firmware -> OS:ACPI->APEI->GHES 
> >  -> CPER -> trace report.  This is the second one.  These two reports  
> > are indicating the same poisoning page, which is the so-called "duplicate
> > report"[1].  And the memory_failure() handling I'm trying to add in
> > OS-First path could also be another duplicate report.
> > 
> > Hope the flow below could make it easier to understand:
> > CPU accesses bad memory on CXL device, then
> >  -> MCE (INT18), *always* report (1)
> >  -> * FW-First (implemented now)
> >       -> CXL device -> FW
> > 	      -> OS:ACPI->APEI->GHES->CPER -> trace report (2.a)  
> >     * OS-First (not implemented yet, I'm working on it)
> >       -> CXL device -> MSI
> > 	      -> OS:CXL driver -> memory_failure() (2.b)  
> > so, the (1) and (2.a/b) are duplicated.
> > 
> > (I didn't get response in my reply for [1] while I have to make patch to
> > solve this problem, so please correct me if my understanding is wrong.)
> > 
> > This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
> > to check whether the current poison page has been reported (if yes,
> > stop the notifier chain, won't call the following memory_failure()
> > to report), into `x86_mce_decoder_chain`.  In this way, if the poison
> > page already handled(recorded and reported) in (1) or (2), the other one
> > won't duplicate the report.  The record could be clear when
> > cxl_clear_poison() is called.
> > 
> > [1] https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/
> > 
> > Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> 
> So poison can be cleared in a number of ways and a CXL poison clear command
> is unfortunately only one of them.  Some architectures have instructions
> that guarantee to write a whole cacheline and can clear things as well.
> I believe x86 does for starters.

Yes, movdir64b.

> +CC linux-edac and related maintainers / reviewers.
>     linux-mm and hwpoison maintainer.
> 
> So I think this needs a more general solution that encompasses 
> more general cleanup of poison.

I think unless the device has "List Poison" coverage for volatile ranges
that the kernel should not worry about tracking this itself.

Perhaps what is needed is that after successful memory_failure()
handling when the page is known to be offline the device backing the
memory can be notified that it is safe to repair the page and but it
back into service, but I expect that would be comparison of the device's
own poison tracking relative to the notification of successful page
offline.

> 
> Trivial comments inline.
> 
> Jonathan
> 
> 
> > ---
> >  arch/x86/include/asm/mce.h |   1 +
> >  drivers/cxl/core/mbox.c    | 130 +++++++++++++++++++++++++++++++++++++
> >  drivers/cxl/core/memdev.c  |   6 +-
> >  drivers/cxl/cxlmem.h       |   3 +
> >  4 files changed, 139 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/x86/include/asm/mce.h b/arch/x86/include/asm/mce.h
> > index dfd2e9699bd7..d8109c48e7d9 100644
> > --- a/arch/x86/include/asm/mce.h
> > +++ b/arch/x86/include/asm/mce.h
> > @@ -182,6 +182,7 @@ enum mce_notifier_prios {
> >  	MCE_PRIO_NFIT,
> >  	MCE_PRIO_EXTLOG,
> >  	MCE_PRIO_UC,
> > +	MCE_PRIO_CXL,
> >  	MCE_PRIO_EARLY,
> >  	MCE_PRIO_CEC,
> >  	MCE_PRIO_HIGHEST = MCE_PRIO_CEC
> > diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> > index 2626f3fff201..0eb3c5401e81 100644
> > --- a/drivers/cxl/core/mbox.c
> > +++ b/drivers/cxl/core/mbox.c
> > @@ -4,6 +4,8 @@
> >  #include <linux/debugfs.h>
> >  #include <linux/ktime.h>
> >  #include <linux/mutex.h>
> > +#include <linux/notifier.h>
> > +#include <asm/mce.h>
> >  #include <asm/unaligned.h>
> >  #include <cxlpci.h>
> >  #include <cxlmem.h>
> > @@ -880,6 +882,9 @@ void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> >  		if (cxlr)
> >  			hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> >  
> > +		if (hpa != ULLONG_MAX && cxl_mce_recorded(hpa))
> > +			return;
> > +
> >  		if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> >  			trace_cxl_general_media(cxlmd, type, cxlr, hpa,
> >  						&evt->gen_media);
> > @@ -1408,6 +1413,127 @@ int cxl_poison_state_init(struct cxl_memdev_state *mds)
> >  }
> >  EXPORT_SYMBOL_NS_GPL(cxl_poison_state_init, CXL);
> >  
> > +struct cxl_mce_record {
> > +	struct list_head node;
> > +	u64 hpa;
> > +};
> > +LIST_HEAD(cxl_mce_records);
> > +DEFINE_MUTEX(cxl_mce_mutex);
> > +
> > +bool cxl_mce_recorded(u64 hpa)
> > +{
> > +	struct cxl_mce_record *cur, *next, *rec;
> > +	int rc;
> > +
> > +	rc = mutex_lock_interruptible(&cxl_mce_mutex);
> 
> guard(mutex)(&cxl_mce_muted);

Agree, _interruptible is really only suitable for user ABI facing locks,
not kernel internal helper functions, but this comment is moot if this
tracking switches to xarray.

> 
> > +	if (rc)
> > +		return false;
> > +
> > +	list_for_each_entry_safe(cur, next, &cxl_mce_records, node) {
> > +		if (cur->hpa == hpa) {
> > +			mutex_unlock(&cxl_mce_mutex);
> > +			return true;
> > +		}
> > +	}
> > +
> > +	rec = kmalloc(sizeof(struct cxl_mce_record), GFP_KERNEL);
> > +	rec->hpa = hpa;
> > +	list_add(&cxl_mce_records, &rec->node);
> > +
> > +	mutex_unlock(&cxl_mce_mutex);
> > +
> > +	return false;
> > +}
> > +
> > +void cxl_mce_clear(u64 hpa)
> > +{
> > +	struct cxl_mce_record *cur, *next;
> > +	int rc;
> > +
> > +	rc = mutex_lock_interruptible(&cxl_mce_mutex);
> 
> Maybe cond_guard().

cond_guard() was rejected, you meant scoped_cond_guard()? But, then I
think _interruptible is not appropriate here.

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Jonathan Cameron]

On Fri, 21 Jun 2024 10:59:46 -0700
Dan Williams <dan.j.williams@intel.com> wrote:

> Jonathan Cameron wrote:
> > On Wed, 19 Jun 2024 00:53:10 +0800
> > Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:
> >   
> > > Background:
> > > Since CXL device is a memory device, while CPU consumes a poison page of 
> > > CXL device, it always triggers a MCE by interrupt (INT18), no matter 
> > > which-First path is configured.  This is the first report.  Then 
> > > currently, in FW-First path, the poison event is transferred according 
> > > to the following process: CXL device -> firmware -> OS:ACPI->APEI->GHES   
> > >  -> CPER -> trace report.  This is the second one.  These two reports    
> > > are indicating the same poisoning page, which is the so-called "duplicate
> > > report"[1].  And the memory_failure() handling I'm trying to add in
> > > OS-First path could also be another duplicate report.
> > > 
> > > Hope the flow below could make it easier to understand:
> > > CPU accesses bad memory on CXL device, then  
> > >  -> MCE (INT18), *always* report (1)
> > >  -> * FW-First (implemented now)
> > >       -> CXL device -> FW
> > > 	      -> OS:ACPI->APEI->GHES->CPER -> trace report (2.a)    
> > >     * OS-First (not implemented yet, I'm working on it)  
> > >       -> CXL device -> MSI
> > > 	      -> OS:CXL driver -> memory_failure() (2.b)    
> > > so, the (1) and (2.a/b) are duplicated.
> > > 
> > > (I didn't get response in my reply for [1] while I have to make patch to
> > > solve this problem, so please correct me if my understanding is wrong.)
> > > 
> > > This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
> > > to check whether the current poison page has been reported (if yes,
> > > stop the notifier chain, won't call the following memory_failure()
> > > to report), into `x86_mce_decoder_chain`.  In this way, if the poison
> > > page already handled(recorded and reported) in (1) or (2), the other one
> > > won't duplicate the report.  The record could be clear when
> > > cxl_clear_poison() is called.
> > > 
> > > [1] https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/
> > > 
> > > Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>  
> > 
> > So poison can be cleared in a number of ways and a CXL poison clear command
> > is unfortunately only one of them.  Some architectures have instructions
> > that guarantee to write a whole cacheline and can clear things as well.
> > I believe x86 does for starters.  
> 
> Yes, movdir64b.

Equivalent arm64 instruction is not valid to normal memory. Lets say
no more on that :(

So who actually cares about recovering poisoned volatile memory?
I'd like to understand more on how significant a use case this is.
Whilst I can conjecture that its an extreme case of wanting to avoid
loosing the ability to create 1GiB or larger pages due to poison
is that a real problem for anyone today?  Note this is just the case
where you've reached an actual uncorrectable error and probably
/ possibly killed something, not the more common soft offlining
of memory due to correctable errors being detected.

> 
> > +CC linux-edac and related maintainers / reviewers.
> >     linux-mm and hwpoison maintainer.
> > 
> > So I think this needs a more general solution that encompasses 
> > more general cleanup of poison.  
> 
> I think unless the device has "List Poison" coverage for volatile ranges
> that the kernel should not worry about tracking this itself.

Maybe.  I think you can still get a media event for this as well
as synchronous poison so there may be a path to a double report, just
a more timely one hopefully.

> 
> Perhaps what is needed is that after successful memory_failure()
> handling when the page is known to be offline the device backing the
> memory can be notified that it is safe to repair the page and but it
> back into service, but I expect that would be comparison of the device's
> own poison tracking relative to the notification of successful page
> offline.

That would work. Elide the error handling if the page is known to
be offline due to poison.  Might be racey though but does it
really hurt if we occasionally report twice?

> > > +	rec = kmalloc(sizeof(struct cxl_mce_record), GFP_KERNEL);
> > > +	rec->hpa = hpa;
> > > +	list_add(&cxl_mce_records, &rec->node);
> > > +
> > > +	mutex_unlock(&cxl_mce_mutex);
> > > +
> > > +	return false;
> > > +}
> > > +
> > > +void cxl_mce_clear(u64 hpa)
> > > +{
> > > +	struct cxl_mce_record *cur, *next;
> > > +	int rc;
> > > +
> > > +	rc = mutex_lock_interruptible(&cxl_mce_mutex);  
> > 
> > Maybe cond_guard().  
> 
> cond_guard() was rejected, you meant scoped_cond_guard()? But, then I
> think _interruptible is not appropriate here.

Ah yes.  Indeed scoped_cond_guard() but fair enough on the
interruptible point!


> 

RE: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Luck, Tony]

> So who actually cares about recovering poisoned volatile memory?
> I'd like to understand more on how significant a use case this is.
> Whilst I can conjecture that its an extreme case of wanting to avoid
> loosing the ability to create 1GiB or larger pages due to poison
> is that a real problem for anyone today?  Note this is just the case
> where you've reached an actual uncorrectable error and probably
> / possibly killed something, not the more common soft offlining
> of memory due to correctable errors being detected.

I guess you really need a reply from someone with a data center
with thousands of machines, since that's where this question
may be important.

My humble opinion is that, outside of the huge page issue, nobody
should try to recover a poisoned page. Systems that can report
and recover from poison have tens, hundreds, or more GBytes
of memory. Dropping 4K pages will not have any measurable
impact on a system (even if there are hundreds of pages dropped).

There's no reliable way to determine whether the poisoned page
was due to some transient issue, or a permanent defect. Recovering
a poisoned page runs the risk that the poison will re-occur. Perhaps
next use of the page will be in some unrecoverable (kernel) context.

So recovery has some risk, but very little upside benefit.

-Tony

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Shiyang Ruan]

在 2024/6/22 4:44, Luck, Tony 写道:
>> So who actually cares about recovering poisoned volatile memory?
>> I'd like to understand more on how significant a use case this is.
>> Whilst I can conjecture that its an extreme case of wanting to avoid
>> loosing the ability to create 1GiB or larger pages due to poison
>> is that a real problem for anyone today?  Note this is just the case
>> where you've reached an actual uncorrectable error and probably
>> / possibly killed something, not the more common soft offlining
>> of memory due to correctable errors being detected.
> 
> I guess you really need a reply from someone with a data center
> with thousands of machines, since that's where this question
> may be important.
> 
> My humble opinion is that, outside of the huge page issue, nobody
> should try to recover a poisoned page. Systems that can report
> and recover from poison have tens, hundreds, or more GBytes
> of memory. Dropping 4K pages will not have any measurable
> impact on a system (even if there are hundreds of pages dropped).
> 
> There's no reliable way to determine whether the poisoned page
> was due to some transient issue, or a permanent defect. Recovering
> a poisoned page runs the risk that the poison will re-occur. Perhaps
> next use of the page will be in some unrecoverable (kernel) context.
> 
> So recovery has some risk, but very little upside benefit.

Since the hardware provides the instruction(CPU)/command(CXL) to clear 
the poison, we could make the function work, at least as an optional 
feature.  Then users could decide to use it or not after evaluating the 
risk and benefit.

I think doing recovery is an improvement step, and may need a lot of 
discussion.  I'm not sure if we could reach a conclusion in this thread. 
  Just hope more comments on the original problem (duplicate report) to 
solve in this patch.


--
Thanks,
Ruan.

> 
> -Tony

RE: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Luck, Tony]

>> So recovery has some risk, but very little upside benefit.
>
> Since the hardware provides the instruction(CPU)/command(CXL) to clear 
> the poison, we could make the function work, at least as an optional 
> feature.  Then users could decide to use it or not after evaluating the 
> risk and benefit.
>
> I think doing recovery is an improvement step, and may need a lot of 
> discussion.  I'm not sure if we could reach a conclusion in this thread. 
>  Just hope more comments on the original problem (duplicate report) to 
> solve in this patch.

Post a patch to implement poison clearing/recovery with some way to opt-in
to get the discussion started. If it doesn't add too much complexity perhaps
it will be accepted.

-Tony

Re: [RFC PATCH] cxl: avoid duplicating report from MCE & device

[Shiyang Ruan]

在 2024/6/19 0:53, Shiyang Ruan 写道:
> 
> This patch adds a new notifier_block and MCE_PRIO_CXL, for CXL memdev
> to check whether the current poison page has been reported (if yes,
> stop the notifier chain, won't call the following memory_failure()
> to report), into `x86_mce_decoder_chain`.  In this way, if the poison
> page already handled(recorded and reported) in (1) or (2), the other one
> won't duplicate the report.  The record could be clear when
> cxl_clear_poison() is called.

Hi guys,

I'd like to sort out the work I am currently carrying forward, to make 
sure I'm not going in the wrong direction. Please correct me if anything 
is wrong.

As is known to us, CXL spec defines POISON feature to notify its status 
when CXL memory device got a broken page.  Basically, there are two 
major paths for the notification.

1. CPU handling error
   When a process is accessing this broken page, CXL device returns data
   with POISON.  When CPU consumes the POISON, it raises a kind of error
   notification.
   To be precise, "how CPU should behave when it consumes POISON" is
   architecture dependent.  In my understanding, x86-64 raises Machine
   Check Exception(MCE) via interrupt #18 in this case.
2. CXL device reporting error
   When CXL device detects the broken page by itself and sends memory
   error signal to kernel in two optional paths.
   2.a. FW-First
     CXL device sends error via VDM to CXL Host, then CXL Host sends it
     to System Firmware via interrupt, finally kernel handles the error.
   2.b. OS-First
     CXL device directly sends error via MSI/MSI-X to kernel.

Note: Since I'm now focusing on x86_64, basically I'll describe about 
x86-64 only.

The following diagram should describe the 2 major paths and 2 optional 
sub-paths above.
```
1.  MCE (interrupt #18, while CPU consuming POISON)
     -> do_machine_check()
       -> mce_log()
         -> notify chain (x86_mce_decoder_chain)
           -> memory_failure()
2.a FW-First (optional, CXL device proactively find&report)
     -> CXL device -> Firmware
       -> OS: ACPI->APEI->GHES->CPER -> CXL driver -> trace
2.b OS-First (optional, CXL device proactively find&report)
     -> CXL device -> MSI
       -> OS: CXL driver -> trace
```

For "1. CPU handling error" path, the current code seems to work fine. 
When I used error injection feature on QEMU emulation, the code path is 
executed certainly.  Then, if the CPU certainly raises a MCE when it 
consumes the POISON, this path has no problem.

So, I'm working on making for 2.a and 2.b path, which is CXL device 
reported POISON error could be handled by kernel.  This path has two 
advantages.

- Proactively find&report memory problems

   Even if a process does not read data yet, kernel/drivers can prevent
   the process from using corrupted data proactively.  AFAIK, the current
   kernel only traces POISON error event from FW-First/OS-First path, but
   it doesn't handle them, neither notify processes who are using the
   POISON page like MCE does.  User space tools like rasdaemon reads the
   trace and log it, but as well, it doesn't handle the POISON page.  As
   a result, user has to read the error log from rasdaemon, distinguish
   whether the POISON error is from CXL memory or DDR memory, find out
   which applications are effected.  That is not an easy work and cannot
   be handled in time.  Thus, I'd like to add a feature to make the work
   done automatically and quickly. Once CXL device reports the POISON
   error (via FW-First/OS-First), kernel handles it immediately, similar
   to the flow when a MCE is triggered.  This is my first motivation.

- Architecture independent

   As the mentioned above, "1. CPU handling error" path is architecture
   dependent.  On the other hand, this route can be architecture
   independent code.  If there is a CPU which does not have similar
   feature like MCE of x86-64, my work will be essential.  (To be honest,
   I did not notice this advantage at first as mentioned later, but I
   think this is also important.)


Here is the timeline of my development of it.

Two series of patches have been sent so far:
- PATCH: cxl/core: add poison creation event handler [1]
- PATCH: cxl: avoid duplicating report from MCE & device [2]
[1] 
https://lore.kernel.org/linux-cxl/20240417075053.3273543-1-ruansy.fnst@fujitsu.com/
[2] 
https://lore.kernel.org/linux-cxl/20240618165310.877974-1-ruansy.fnst@fujitsu.com/


The 1st patch[1] added POISON error handler in "2. CXL device reporting 
error" path.

My first version was constructing a MCE data from POISON address and 
calling mce_log() to handle the POISON.  But I was told that 
constructing MCE data is architecture dependent while CXL is not.  So, 
in later version, just call memory_failure_queue() in CXL to handle the 
POISON error to avoid the arch-dependent problem.

After many discussions, a new problem was found: as Dan said[3], added 
POISON handling will cause the "duplicate report" problem:
 > So, I think all CXL poison notification events should trigger an
 > action optional memory_failure(). I expect this needs to make sure
 > that duplicates re not a problem. I.e. in the case of CPU consumption
 > of CXL poison, that causes a synchronous MF_ACTION_REQUIRED event via
 > the MCE path *and* it may trigger the device to send an error record
 > for the same page. As far as I can see, duplicate reports (MCE + CXL
 > device) are unavoidable.

[3] 
https://lore.kernel.org/linux-cxl/664d948fb86f0_e8be294f8@dwillia2-mobl3.amr.corp.intel.com.notmuch/


To solve this problem, I made the 2nd patch[2].  Allow me to describe 
the background again:

Since CXL device is a memory device, while CPU is consuming a poison 
page of CXL device, it always triggers a MCE (via interrupt #18) and 
calls memory_failure() to handle POISON page, no matter which-First path 
is configured.

My patch added memory_failure() in FW-First/OS-First path: if device 
finds and reports the POISON, kernel not only traces but also calls 
memory_failure() to handle it, marked as "ADD" in the figure blow.
```
1.  MCE (interrupt #18, while CPU consuming POISON)
     -> do_machine_check()
       -> mce_log()
         -> notify chain (x86_mce_decoder_chain)
           -> memory_failure() <---------------------------- EXISTS
2.a FW-First (optional, CXL device proactively find&report)
     -> CXL device -> Firmware
       -> OS: ACPI->APEI->GHES->CPER -> CXL driver -> trace
                                                  \-> memory_failure()
                                                      ^----- ADD
2.b OS-First (optional, CXL device proactively find&report)
     -> CXL device -> MSI
       -> OS: CXL driver -> trace
                        \-> memory_failure()
                            ^------------------------------- ADD
```

But in this way, the memory_failure() could be called twice or even at 
same time, as is shown in the figure above: (1.) and (2.a or 2.b), 
before the POISON page is cleared.  memory_failure() has it own mutex 
lock so it actually won't be called at same time and the later call 
could be avoided because HWPoison bit has been set.  However, assume 
such a scenario, "CXL device reports POISON error" triggers 1st call, 
user see it from log and want to clear the poison by executing `cxl 
clear-poison` command, and at the same time, a process tries to access 
this POISON page, which triggers MCE (it's the 2nd call).  Since there 
is no lock between the 2nd call with clearing poison operation, race 
condition may happen, which may cause HWPoison bit of the page in an 
unknown state.

Thus, we have to avoid the 2nd call. This patch[2] introduces a new 
notifier_block into `x86_mce_decoder_chain` and a POISON cache list, to 
stop the 2nd call of memory_failure(). It checks whether the current 
poison page has been reported (if yes, stop the notifier chain, don't 
call the following memory_failure() to report again).

Looking forward to your comments!


--
Thanks,
Ruan.