diff mbox series

[v3,2/2] cxl/core: add poison creation event handler

Message ID 20240417075053.3273543-3-ruansy.fnst@fujitsu.com
State New, archived
Headers show
Series cxl: add poison creation event handler | expand

Commit Message

Shiyang Ruan April 17, 2024, 7:50 a.m. UTC
Currently driver only traces cxl events, poison creation (for both vmem
and pmem type) on cxl memdev is silent.  OS needs to be notified then it
could handle poison pages in time.  Per CXL spec, the device error event
could be signaled through FW-First and OS-First methods.

So, add poison creation event handler in OS-First method:
  - Qemu:
    - CXL device reports POISON creation event to OS by MSI by sending
      GMER/DER after injecting a poison record;
  - CXL driver:
    a. parse the POISON event from GMER/DER;
    b. translate poisoned DPA to HPA (PFN);
    c. enqueue poisoned PFN to memory_failure's work queue;

Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
---
 drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
 drivers/cxl/cxlmem.h      |   8 +--
 include/linux/cxl-event.h |  18 +++++-
 3 files changed, 125 insertions(+), 20 deletions(-)

Comments

Dave Jiang April 17, 2024, 5:30 p.m. UTC | #1
On 4/17/24 12:50 AM, Shiyang Ruan wrote:
> Currently driver only traces cxl events, poison creation (for both vmem
> and pmem type) on cxl memdev is silent.  OS needs to be notified then it
> could handle poison pages in time.  Per CXL spec, the device error event
> could be signaled through FW-First and OS-First methods.

Please consider below for better clarity:
Currently the driver only traces CXL events. Poison creation (for both ram
and pmem type) on a CXL memdev is silent. The OS needs to be notified so it
can handle poison pages. Per CXL spec, the device error event
can be signaled through the FW-First method or the OS-First method.

> 
> So, add poison creation event handler in OS-First method:
>   - Qemu:
>     - CXL device reports POISON creation event to OS by MSI by sending
>       GMER/DER after injecting a poison record;

Can probably drop the QEMU changes and this is the kernel commit log.

>   - CXL driver:
>     a. parse the POISON event from GMER/DER;
>     b. translate poisoned DPA to HPA (PFN);
>     c. enqueue poisoned PFN to memory_failure's work queue;
> 
> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> ---
>  drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>  drivers/cxl/cxlmem.h      |   8 +--
>  include/linux/cxl-event.h |  18 +++++-
>  3 files changed, 125 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> index f0f54aeccc87..76af0d73859d 100644
> --- a/drivers/cxl/core/mbox.c
> +++ b/drivers/cxl/core/mbox.c
> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>  }
>  EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>  
> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -			    enum cxl_event_log_type type,
> -			    enum cxl_event_type event_type,
> -			    const uuid_t *uuid, union cxl_event *evt)
> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,

I think this needs to be changed to __cxl_report_poison() and the function below to cxl_report_poison(). Otherwise it goes against typical Linux methodology of having the __functionX() as the raw functionality function called by a functionX() wrapper. 

DJ

> +			      u64 dpa)
>  {
> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> +	unsigned long pfn = PHYS_PFN(hpa);
> +
> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
> +		return;
> +
> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);
> +}
> +
> +static int __cxl_report_poison(struct device *dev, void *arg)
> +{
> +	struct cxl_endpoint_decoder *cxled;
> +	struct cxl_memdev *cxlmd;
> +	u64 dpa = *(u64 *)arg;
> +
> +	cxled = to_cxl_endpoint_decoder(dev);
> +	if (!cxled || !cxled->dpa_res || !resource_size(cxled->dpa_res))
> +		return 0;
> +
> +	if (cxled->mode == CXL_DECODER_MIXED) {
> +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> +		return 0;
> +	}
> +
> +	if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
> +		return 0;
> +
> +	cxlmd = cxled_to_memdev(cxled);
> +	cxl_report_poison(cxlmd, cxled->cxld.region, dpa);
> +
> +	return 1;
> +}
> +
> +static void cxl_event_handle_poison(struct cxl_memdev *cxlmd, u64 dpa)
> +{
> +	struct cxl_port *port = cxlmd->endpoint;
> +
> +	/*
> +	 * No region is mapped to this endpoint, that is to say no HPA is
> +	 * mapped.
> +	 */
> +	if (!port || !is_cxl_endpoint(port) ||
> +	    cxl_num_decoders_committed(port) == 0)
> +		return;
> +
> +	device_for_each_child(&port->dev, &dpa, __cxl_report_poison);
> +}
> +
> +static void cxl_event_handle_general_media(struct cxl_memdev *cxlmd,
> +					   enum cxl_event_log_type type,
> +					   struct cxl_event_gen_media *rec)
> +{
> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
> +
> +	if (type == CXL_EVENT_TYPE_FAIL) {
> +		switch (rec->transaction_type) {
> +		case CXL_EVENT_TRANSACTION_READ:
> +		case CXL_EVENT_TRANSACTION_WRITE:
> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:
> +			cxl_event_handle_poison(cxlmd, dpa);
> +			break;
> +		default:
> +			break;
> +		}
> +	}
> +}
> +
> +static void cxl_event_handle_dram(struct cxl_memdev *cxlmd,
> +				  enum cxl_event_log_type type,
> +				  struct cxl_event_dram *rec)
> +{
> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
> +
> +	if (type == CXL_EVENT_TYPE_FAIL) {
> +		switch (rec->transaction_type) {
> +		case CXL_EVENT_TRANSACTION_READ:
> +		case CXL_EVENT_TRANSACTION_WRITE:
> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:
> +			cxl_event_handle_poison(cxlmd, dpa);
> +			break;
> +		default:
> +			break;
> +		}
> +	}
> +}
> +
> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +			     enum cxl_event_log_type type,
> +			     enum cxl_event_type event_type,
> +			     const uuid_t *uuid, union cxl_event *evt)
> +{
> +	if (event_type == CXL_CPER_EVENT_GEN_MEDIA) {
>  		trace_cxl_general_media(cxlmd, type, &evt->gen_media);
> -	else if (event_type == CXL_CPER_EVENT_DRAM)
> +		cxl_event_handle_general_media(cxlmd, type, &evt->gen_media);
> +	} else if (event_type == CXL_CPER_EVENT_DRAM) {
>  		trace_cxl_dram(cxlmd, type, &evt->dram);
> -	else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
> +		cxl_event_handle_dram(cxlmd, type, &evt->dram);
> +	} else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
>  		trace_cxl_memory_module(cxlmd, type, &evt->mem_module);
>  	else
>  		trace_cxl_generic_event(cxlmd, type, uuid, &evt->generic);
>  }
> -EXPORT_SYMBOL_NS_GPL(cxl_event_trace_record, CXL);
> +EXPORT_SYMBOL_NS_GPL(cxl_event_handle_record, CXL);
>  
> -static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -				     enum cxl_event_log_type type,
> -				     struct cxl_event_record_raw *record)
> +static void __cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +				      enum cxl_event_log_type type,
> +				      struct cxl_event_record_raw *record)
>  {
>  	enum cxl_event_type ev_type = CXL_CPER_EVENT_GENERIC;
>  	const uuid_t *uuid = &record->id;
> @@ -867,7 +958,7 @@ static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>  	else if (uuid_equal(uuid, &CXL_EVENT_MEM_MODULE_UUID))
>  		ev_type = CXL_CPER_EVENT_MEM_MODULE;
>  
> -	cxl_event_trace_record(cxlmd, type, ev_type, uuid, &record->event);
> +	cxl_event_handle_record(cxlmd, type, ev_type, uuid, &record->event);
>  }
>  
>  static int cxl_clear_event_record(struct cxl_memdev_state *mds,
> @@ -979,8 +1070,8 @@ static void cxl_mem_get_records_log(struct cxl_memdev_state *mds,
>  			break;
>  
>  		for (i = 0; i < nr_rec; i++)
> -			__cxl_event_trace_record(cxlmd, type,
> -						 &payload->records[i]);
> +			__cxl_event_handle_record(cxlmd, type,
> +						  &payload->records[i]);
>  
>  		if (payload->flags & CXL_GET_EVENT_FLAG_OVERFLOW)
>  			trace_cxl_overflow(cxlmd, type, payload);
> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
> index 36cee9c30ceb..ba1347de5651 100644
> --- a/drivers/cxl/cxlmem.h
> +++ b/drivers/cxl/cxlmem.h
> @@ -822,10 +822,10 @@ void set_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>  void clear_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>  				  unsigned long *cmds);
>  void cxl_mem_get_event_records(struct cxl_memdev_state *mds, u32 status);
> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -			    enum cxl_event_log_type type,
> -			    enum cxl_event_type event_type,
> -			    const uuid_t *uuid, union cxl_event *evt);
> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +			     enum cxl_event_log_type type,
> +			     enum cxl_event_type event_type,
> +			     const uuid_t *uuid, union cxl_event *evt);
>  int cxl_set_timestamp(struct cxl_memdev_state *mds);
>  int cxl_poison_state_init(struct cxl_memdev_state *mds);
>  int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
> diff --git a/include/linux/cxl-event.h b/include/linux/cxl-event.h
> index 03fa6d50d46f..8189bed76c12 100644
> --- a/include/linux/cxl-event.h
> +++ b/include/linux/cxl-event.h
> @@ -23,6 +23,20 @@ struct cxl_event_generic {
>  	u8 data[CXL_EVENT_RECORD_DATA_LENGTH];
>  } __packed;
>  
> +/*
> + * Event transaction type
> + * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
> + */
> +enum cxl_event_transaction_type {
> +	CXL_EVENT_TRANSACTION_UNKNOWN = 0X00,
> +	CXL_EVENT_TRANSACTION_READ,
> +	CXL_EVENT_TRANSACTION_WRITE,
> +	CXL_EVENT_TRANSACTION_SCAN_MEDIA,
> +	CXL_EVENT_TRANSACTION_INJECT_POISON,
> +	CXL_EVENT_TRANSACTION_MEDIA_SCRUB,
> +	CXL_EVENT_TRANSACTION_MEDIA_MANAGEMENT,
> +};
> +
>  /*
>   * General Media Event Record
>   * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
> @@ -33,7 +47,7 @@ struct cxl_event_gen_media {
>  	__le64 phys_addr;
>  	u8 descriptor;
>  	u8 type;
> -	u8 transaction_type;
> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>  	u8 validity_flags[2];
>  	u8 channel;
>  	u8 rank;
> @@ -52,7 +66,7 @@ struct cxl_event_dram {
>  	__le64 phys_addr;
>  	u8 descriptor;
>  	u8 type;
> -	u8 transaction_type;
> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>  	u8 validity_flags[2];
>  	u8 channel;
>  	u8 rank;
Shiyang Ruan April 18, 2024, 9:01 a.m. UTC | #2
在 2024/4/18 1:30, Dave Jiang 写道:
> 
> 
> On 4/17/24 12:50 AM, Shiyang Ruan wrote:
>> Currently driver only traces cxl events, poison creation (for both vmem
>> and pmem type) on cxl memdev is silent.  OS needs to be notified then it
>> could handle poison pages in time.  Per CXL spec, the device error event
>> could be signaled through FW-First and OS-First methods.
> 
> Please consider below for better clarity:
> Currently the driver only traces CXL events. Poison creation (for both ram
> and pmem type) on a CXL memdev is silent. The OS needs to be notified so it
> can handle poison pages. Per CXL spec, the device error event
> can be signaled through the FW-First method or the OS-First method.

Thanks, this is better.

> 
>>
>> So, add poison creation event handler in OS-First method:
>>    - Qemu:
>>      - CXL device reports POISON creation event to OS by MSI by sending
>>        GMER/DER after injecting a poison record;
> 
> Can probably drop the QEMU changes and this is the kernel commit log.

Ok.

> 
>>    - CXL driver:
>>      a. parse the POISON event from GMER/DER;
>>      b. translate poisoned DPA to HPA (PFN);
>>      c. enqueue poisoned PFN to memory_failure's work queue;
>>
>> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
>> ---
>>   drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>>   drivers/cxl/cxlmem.h      |   8 +--
>>   include/linux/cxl-event.h |  18 +++++-
>>   3 files changed, 125 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
>> index f0f54aeccc87..76af0d73859d 100644
>> --- a/drivers/cxl/core/mbox.c
>> +++ b/drivers/cxl/core/mbox.c
>> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>>   }
>>   EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>>   
>> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>> -			    enum cxl_event_log_type type,
>> -			    enum cxl_event_type event_type,
>> -			    const uuid_t *uuid, union cxl_event *evt)
>> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
> 
> I think this needs to be changed to __cxl_report_poison() and the function below to cxl_report_poison(). Otherwise it goes against typical Linux methodology of having the __functionX() as the raw functionality function called by a functionX() wrapper.

This function was designed to do the real reporting work, and could be 
called at other places (actually did in previous version).  Now that it 
is called only below in this version, yes, it's better to change the names.


--
Thanks,
Ruan.

> 
> DJ
>
kernel test robot April 21, 2024, 12:14 p.m. UTC | #3
Hi Shiyang,

kernel test robot noticed the following build errors:

[auto build test ERROR on cxl/next]
[also build test ERROR on linus/master v6.9-rc4 next-20240419]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Shiyang-Ruan/cxl-core-correct-length-of-DPA-field-masks/20240417-155443
base:   https://git.kernel.org/pub/scm/linux/kernel/git/cxl/cxl.git next
patch link:    https://lore.kernel.org/r/20240417075053.3273543-3-ruansy.fnst%40fujitsu.com
patch subject: [PATCH v3 2/2] cxl/core: add poison creation event handler
config: csky-randconfig-002-20240421 (https://download.01.org/0day-ci/archive/20240421/202404212044.uSxtGRtL-lkp@intel.com/config)
compiler: csky-linux-gcc (GCC) 13.2.0
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20240421/202404212044.uSxtGRtL-lkp@intel.com/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <lkp@intel.com>
| Closes: https://lore.kernel.org/oe-kbuild-all/202404212044.uSxtGRtL-lkp@intel.com/

All errors (new ones prefixed by >>):

   `.exit.text' referenced in section `__jump_table' of fs/ceph/super.o: defined in discarded section `.exit.text' of fs/ceph/super.o
   `.exit.text' referenced in section `__jump_table' of fs/ceph/super.o: defined in discarded section `.exit.text' of fs/ceph/super.o
   csky-linux-ld: drivers/cxl/core/mbox.o: in function `__cxl_report_poison':
   mbox.c:(.text+0xa12): undefined reference to `cxl_trace_hpa'
>> csky-linux-ld: mbox.c:(.text+0xa44): undefined reference to `cxl_trace_hpa'
   `.exit.text' referenced in section `__jump_table' of drivers/target/target_core_configfs.o: defined in discarded section `.exit.text' of drivers/target/target_core_configfs.o
   `.exit.text' referenced in section `__jump_table' of drivers/target/target_core_configfs.o: defined in discarded section `.exit.text' of drivers/target/target_core_configfs.o
   `.exit.text' referenced in section `__jump_table' of net/ceph/ceph_common.o: defined in discarded section `.exit.text' of net/ceph/ceph_common.o
   `.exit.text' referenced in section `__jump_table' of net/ceph/ceph_common.o: defined in discarded section `.exit.text' of net/ceph/ceph_common.o
Ira Weiny April 23, 2024, 5:57 p.m. UTC | #4
Shiyang Ruan wrote:
> Currently driver only traces cxl events, poison creation (for both vmem
> and pmem type) on cxl memdev is silent.  OS needs to be notified then it
> could handle poison pages in time.  Per CXL spec, the device error event
> could be signaled through FW-First and OS-First methods.
> 
> So, add poison creation event handler in OS-First method:
>   - Qemu:
>     - CXL device reports POISON creation event to OS by MSI by sending
>       GMER/DER after injecting a poison record;
>   - CXL driver:
>     a. parse the POISON event from GMER/DER;
>     b. translate poisoned DPA to HPA (PFN);
>     c. enqueue poisoned PFN to memory_failure's work queue;

I'm a bit confused by the need for this patch.  Perhaps a bit more detail
here?

More comments below.

> 
> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> ---
>  drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>  drivers/cxl/cxlmem.h      |   8 +--
>  include/linux/cxl-event.h |  18 +++++-
>  3 files changed, 125 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> index f0f54aeccc87..76af0d73859d 100644
> --- a/drivers/cxl/core/mbox.c
> +++ b/drivers/cxl/core/mbox.c
> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>  }
>  EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>  
> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -			    enum cxl_event_log_type type,
> -			    enum cxl_event_type event_type,
> -			    const uuid_t *uuid, union cxl_event *evt)
> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
> +			      u64 dpa)
>  {
> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> +	unsigned long pfn = PHYS_PFN(hpa);
> +
> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
> +		return;
> +
> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);

I thought that ras daemon was supposed to take care of this when the trace
event occurred.  Alison is working on the HPA data for that path.

> +}
> +
> +static int __cxl_report_poison(struct device *dev, void *arg)
> +{
> +	struct cxl_endpoint_decoder *cxled;
> +	struct cxl_memdev *cxlmd;
> +	u64 dpa = *(u64 *)arg;
> +
> +	cxled = to_cxl_endpoint_decoder(dev);
> +	if (!cxled || !cxled->dpa_res || !resource_size(cxled->dpa_res))
> +		return 0;
> +
> +	if (cxled->mode == CXL_DECODER_MIXED) {
> +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> +		return 0;
> +	}
> +
> +	if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
> +		return 0;
> +
> +	cxlmd = cxled_to_memdev(cxled);
> +	cxl_report_poison(cxlmd, cxled->cxld.region, dpa);
> +
> +	return 1;
> +}
> +
> +static void cxl_event_handle_poison(struct cxl_memdev *cxlmd, u64 dpa)
> +{
> +	struct cxl_port *port = cxlmd->endpoint;
> +
> +	/*
> +	 * No region is mapped to this endpoint, that is to say no HPA is
> +	 * mapped.
> +	 */
> +	if (!port || !is_cxl_endpoint(port) ||
> +	    cxl_num_decoders_committed(port) == 0)
> +		return;
> +
> +	device_for_each_child(&port->dev, &dpa, __cxl_report_poison);
> +}
> +
> +static void cxl_event_handle_general_media(struct cxl_memdev *cxlmd,
> +					   enum cxl_event_log_type type,
> +					   struct cxl_event_gen_media *rec)
> +{
> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
> +
> +	if (type == CXL_EVENT_TYPE_FAIL) {

Why only FAIL and not FATAL?

> +		switch (rec->transaction_type) {
> +		case CXL_EVENT_TRANSACTION_READ:
> +		case CXL_EVENT_TRANSACTION_WRITE:
> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:

Why not scan media?

> +			cxl_event_handle_poison(cxlmd, dpa);
> +			break;
> +		default:
> +			break;
> +		}
> +	}
> +}
> +
> +static void cxl_event_handle_dram(struct cxl_memdev *cxlmd,
> +				  enum cxl_event_log_type type,
> +				  struct cxl_event_dram *rec)
> +{
> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
> +
> +	if (type == CXL_EVENT_TYPE_FAIL) {
> +		switch (rec->transaction_type) {
> +		case CXL_EVENT_TRANSACTION_READ:
> +		case CXL_EVENT_TRANSACTION_WRITE:
> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:
> +			cxl_event_handle_poison(cxlmd, dpa);
> +			break;
> +		default:
> +			break;
> +		}
> +	}
> +}
> +
> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +			     enum cxl_event_log_type type,
> +			     enum cxl_event_type event_type,
> +			     const uuid_t *uuid, union cxl_event *evt)
> +{
> +	if (event_type == CXL_CPER_EVENT_GEN_MEDIA) {
>  		trace_cxl_general_media(cxlmd, type, &evt->gen_media);
> -	else if (event_type == CXL_CPER_EVENT_DRAM)
> +		cxl_event_handle_general_media(cxlmd, type, &evt->gen_media);
> +	} else if (event_type == CXL_CPER_EVENT_DRAM) {
>  		trace_cxl_dram(cxlmd, type, &evt->dram);
> -	else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
> +		cxl_event_handle_dram(cxlmd, type, &evt->dram);
> +	} else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
>  		trace_cxl_memory_module(cxlmd, type, &evt->mem_module);
>  	else
>  		trace_cxl_generic_event(cxlmd, type, uuid, &evt->generic);
>  }
> -EXPORT_SYMBOL_NS_GPL(cxl_event_trace_record, CXL);
> +EXPORT_SYMBOL_NS_GPL(cxl_event_handle_record, CXL);
>

Why all the churn with changing the names of functions?

Ira

>  
> -static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -				     enum cxl_event_log_type type,
> -				     struct cxl_event_record_raw *record)
> +static void __cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +				      enum cxl_event_log_type type,
> +				      struct cxl_event_record_raw *record)
>  {
>  	enum cxl_event_type ev_type = CXL_CPER_EVENT_GENERIC;
>  	const uuid_t *uuid = &record->id;
> @@ -867,7 +958,7 @@ static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>  	else if (uuid_equal(uuid, &CXL_EVENT_MEM_MODULE_UUID))
>  		ev_type = CXL_CPER_EVENT_MEM_MODULE;
>  
> -	cxl_event_trace_record(cxlmd, type, ev_type, uuid, &record->event);
> +	cxl_event_handle_record(cxlmd, type, ev_type, uuid, &record->event);
>  }
>  
>  static int cxl_clear_event_record(struct cxl_memdev_state *mds,
> @@ -979,8 +1070,8 @@ static void cxl_mem_get_records_log(struct cxl_memdev_state *mds,
>  			break;
>  
>  		for (i = 0; i < nr_rec; i++)
> -			__cxl_event_trace_record(cxlmd, type,
> -						 &payload->records[i]);
> +			__cxl_event_handle_record(cxlmd, type,
> +						  &payload->records[i]);
>  
>  		if (payload->flags & CXL_GET_EVENT_FLAG_OVERFLOW)
>  			trace_cxl_overflow(cxlmd, type, payload);
> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
> index 36cee9c30ceb..ba1347de5651 100644
> --- a/drivers/cxl/cxlmem.h
> +++ b/drivers/cxl/cxlmem.h
> @@ -822,10 +822,10 @@ void set_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>  void clear_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>  				  unsigned long *cmds);
>  void cxl_mem_get_event_records(struct cxl_memdev_state *mds, u32 status);
> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -			    enum cxl_event_log_type type,
> -			    enum cxl_event_type event_type,
> -			    const uuid_t *uuid, union cxl_event *evt);
> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
> +			     enum cxl_event_log_type type,
> +			     enum cxl_event_type event_type,
> +			     const uuid_t *uuid, union cxl_event *evt);
>  int cxl_set_timestamp(struct cxl_memdev_state *mds);
>  int cxl_poison_state_init(struct cxl_memdev_state *mds);
>  int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
> diff --git a/include/linux/cxl-event.h b/include/linux/cxl-event.h
> index 03fa6d50d46f..8189bed76c12 100644
> --- a/include/linux/cxl-event.h
> +++ b/include/linux/cxl-event.h
> @@ -23,6 +23,20 @@ struct cxl_event_generic {
>  	u8 data[CXL_EVENT_RECORD_DATA_LENGTH];
>  } __packed;
>  
> +/*
> + * Event transaction type
> + * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
> + */
> +enum cxl_event_transaction_type {
> +	CXL_EVENT_TRANSACTION_UNKNOWN = 0X00,
> +	CXL_EVENT_TRANSACTION_READ,
> +	CXL_EVENT_TRANSACTION_WRITE,
> +	CXL_EVENT_TRANSACTION_SCAN_MEDIA,
> +	CXL_EVENT_TRANSACTION_INJECT_POISON,
> +	CXL_EVENT_TRANSACTION_MEDIA_SCRUB,
> +	CXL_EVENT_TRANSACTION_MEDIA_MANAGEMENT,
> +};
> +
>  /*
>   * General Media Event Record
>   * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
> @@ -33,7 +47,7 @@ struct cxl_event_gen_media {
>  	__le64 phys_addr;
>  	u8 descriptor;
>  	u8 type;
> -	u8 transaction_type;
> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>  	u8 validity_flags[2];
>  	u8 channel;
>  	u8 rank;
> @@ -52,7 +66,7 @@ struct cxl_event_dram {
>  	__le64 phys_addr;
>  	u8 descriptor;
>  	u8 type;
> -	u8 transaction_type;
> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>  	u8 validity_flags[2];
>  	u8 channel;
>  	u8 rank;
> -- 
> 2.34.1
>
Dan Williams April 23, 2024, 6:40 p.m. UTC | #5
Shiyang Ruan wrote:
> Currently driver only traces cxl events, poison creation (for both vmem
> and pmem type) on cxl memdev is silent.

As it should be.

> OS needs to be notified then it could handle poison pages in time.

No, it was always the case that latent poison is an "action optional"
event. I am not understanding the justification for this approach. What
breaks if the kernel does not forward events to memory_failure_queue()?

Consider that in the CPU consumption case that the firmware first path
will do its own memory_failure_queue() and in the native case the MCE
handler will take care of this. So that leaves pages that are accessed
by DMA or background operation that encounter poison. Those are "action
optional" scenarios and it is not clear to me how the driver tells the
difference.

This needs more precision on which agent is repsonsible for what level
of reporting. The distribution of responsibility between ACPI GHES,
EDAC, and the CXL driver is messy and I expect this changelog to
demonstrate it understands all those considerations.

> Per CXL spec, the device error event could be signaled through
> FW-First and OS-First methods.
> 
> So, add poison creation event handler in OS-First method:
>   - Qemu:

Why is QEMU relevant for this patch? QEMU is only a development vehicle
the upstream enabling should be reference shipping or expected to be
shipping hardware implementations.

>     - CXL device reports POISON creation event to OS by MSI by sending
>       GMER/DER after injecting a poison record;

When you say "inject" here do you mean "add to the poison list if
present". Because "inject" to me means the "Inject Poison" Memory Device
Command.

>   - CXL driver:
>     a. parse the POISON event from GMER/DER;
>     b. translate poisoned DPA to HPA (PFN);
>     c. enqueue poisoned PFN to memory_failure's work queue;
> 
> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> ---
>  drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>  drivers/cxl/cxlmem.h      |   8 +--
>  include/linux/cxl-event.h |  18 +++++-
>  3 files changed, 125 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> index f0f54aeccc87..76af0d73859d 100644
> --- a/drivers/cxl/core/mbox.c
> +++ b/drivers/cxl/core/mbox.c
> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>  }
>  EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>  
> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> -			    enum cxl_event_log_type type,
> -			    enum cxl_event_type event_type,
> -			    const uuid_t *uuid, union cxl_event *evt)
> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
> +			      u64 dpa)
>  {
> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> +	unsigned long pfn = PHYS_PFN(hpa);
> +
> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
> +		return;

No need for this check, memory_failure_queue() is already stubbed out in
the CONFIG_MEMORY_FAILURE=n case.

> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);

My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
reported errors since action is only required for direct consumption
events and those need not be reported through the device event queue.

It would be useful to collaborate with a BIOS firmware engineer so that
the kernel ends up with similar logic as is used to set CPER record
severity, or at least understands why it would want to be different.
See how ghes_handle_memory_failure() determines the
memory_failure_queue() flags.
Shiyang Ruan May 3, 2024, 10:42 a.m. UTC | #6
在 2024/4/24 1:57, Ira Weiny 写道:
> Shiyang Ruan wrote:
>> Currently driver only traces cxl events, poison creation (for both vmem
>> and pmem type) on cxl memdev is silent.  OS needs to be notified then it
>> could handle poison pages in time.  Per CXL spec, the device error event
>> could be signaled through FW-First and OS-First methods.
>>
>> So, add poison creation event handler in OS-First method:
>>    - Qemu:
>>      - CXL device reports POISON creation event to OS by MSI by sending
>>        GMER/DER after injecting a poison record;
>>    - CXL driver:
>>      a. parse the POISON event from GMER/DER;
>>      b. translate poisoned DPA to HPA (PFN);
>>      c. enqueue poisoned PFN to memory_failure's work queue;
> 
> I'm a bit confused by the need for this patch.  Perhaps a bit more detail
> here?

Yes, I should have wrote more details.

I want to check and make sure the HWPOISON on a CXL device (type3) is 
working properly.  For example, a FSDAX filesystem created on a 
type3-pmem device, then it gets a POISON bit, the OS should be able to 
handle this POISON event: find the relevant process

Currently I'm using Qemu with several simulated CXL devices, and using 
poison injection API of Qemu to create POISON records, but OS isn't 
notified.  Only when we actively call list POISON records (cxl list -L) 
will the driver fetch them and log into trace events, then we see the 
POISON records.  Memory failure wasn't triggered too.

That's why I said "poison creation on cxl memdev is silent".  Per spec, 
POISON creation should be notified to OS.  Since not familiar with 
firmware part, I'm try adding this notification for OS-First.

> 
> More comments below.
> 
>>
>> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
>> ---
>>   drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>>   drivers/cxl/cxlmem.h      |   8 +--
>>   include/linux/cxl-event.h |  18 +++++-
>>   3 files changed, 125 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
>> index f0f54aeccc87..76af0d73859d 100644
>> --- a/drivers/cxl/core/mbox.c
>> +++ b/drivers/cxl/core/mbox.c
>> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>>   }
>>   EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>>   
>> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>> -			    enum cxl_event_log_type type,
>> -			    enum cxl_event_type event_type,
>> -			    const uuid_t *uuid, union cxl_event *evt)
>> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
>> +			      u64 dpa)
>>   {
>> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
>> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>> +	unsigned long pfn = PHYS_PFN(hpa);
>> +
>> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
>> +		return;
>> +
>> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);
> 
> I thought that ras daemon was supposed to take care of this when the trace
> event occurred.  Alison is working on the HPA data for that path.

It seems to save CXL trace events/memory-failures to DB and report to 
others, but cannot let OS call memory failure.

> 
>> +}
>> +
>> +static int __cxl_report_poison(struct device *dev, void *arg)
>> +{
>> +	struct cxl_endpoint_decoder *cxled;
>> +	struct cxl_memdev *cxlmd;
>> +	u64 dpa = *(u64 *)arg;
>> +
>> +	cxled = to_cxl_endpoint_decoder(dev);
>> +	if (!cxled || !cxled->dpa_res || !resource_size(cxled->dpa_res))
>> +		return 0;
>> +
>> +	if (cxled->mode == CXL_DECODER_MIXED) {
>> +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
>> +		return 0;
>> +	}
>> +
>> +	if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
>> +		return 0;
>> +
>> +	cxlmd = cxled_to_memdev(cxled);
>> +	cxl_report_poison(cxlmd, cxled->cxld.region, dpa);
>> +
>> +	return 1;
>> +}
>> +
>> +static void cxl_event_handle_poison(struct cxl_memdev *cxlmd, u64 dpa)
>> +{
>> +	struct cxl_port *port = cxlmd->endpoint;
>> +
>> +	/*
>> +	 * No region is mapped to this endpoint, that is to say no HPA is
>> +	 * mapped.
>> +	 */
>> +	if (!port || !is_cxl_endpoint(port) ||
>> +	    cxl_num_decoders_committed(port) == 0)
>> +		return;
>> +
>> +	device_for_each_child(&port->dev, &dpa, __cxl_report_poison);
>> +}
>> +
>> +static void cxl_event_handle_general_media(struct cxl_memdev *cxlmd,
>> +					   enum cxl_event_log_type type,
>> +					   struct cxl_event_gen_media *rec)
>> +{
>> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
>> +
>> +	if (type == CXL_EVENT_TYPE_FAIL) {
> 
> Why only FAIL and not FATAL?

POISON is recoverable(could be cleared by sending CLEAR_POISON command) 
and won't cause serious problems, so only use failure event here.

> 
>> +		switch (rec->transaction_type) {
>> +		case CXL_EVENT_TRANSACTION_READ:
>> +		case CXL_EVENT_TRANSACTION_WRITE:
>> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:
> 
> Why not scan media?

Yes, this type should be handled too.  Will add.

> 
>> +			cxl_event_handle_poison(cxlmd, dpa);
>> +			break;
>> +		default:
>> +			break;
>> +		}
>> +	}
>> +}
>> +
>> +static void cxl_event_handle_dram(struct cxl_memdev *cxlmd,
>> +				  enum cxl_event_log_type type,
>> +				  struct cxl_event_dram *rec)
>> +{
>> +	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
>> +
>> +	if (type == CXL_EVENT_TYPE_FAIL) {
>> +		switch (rec->transaction_type) {
>> +		case CXL_EVENT_TRANSACTION_READ:
>> +		case CXL_EVENT_TRANSACTION_WRITE:
>> +		case CXL_EVENT_TRANSACTION_INJECT_POISON:
>> +			cxl_event_handle_poison(cxlmd, dpa);
>> +			break;
>> +		default:
>> +			break;
>> +		}
>> +	}
>> +}
>> +
>> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
>> +			     enum cxl_event_log_type type,
>> +			     enum cxl_event_type event_type,
>> +			     const uuid_t *uuid, union cxl_event *evt)
>> +{
>> +	if (event_type == CXL_CPER_EVENT_GEN_MEDIA) {
>>   		trace_cxl_general_media(cxlmd, type, &evt->gen_media);
>> -	else if (event_type == CXL_CPER_EVENT_DRAM)
>> +		cxl_event_handle_general_media(cxlmd, type, &evt->gen_media);
>> +	} else if (event_type == CXL_CPER_EVENT_DRAM) {
>>   		trace_cxl_dram(cxlmd, type, &evt->dram);
>> -	else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
>> +		cxl_event_handle_dram(cxlmd, type, &evt->dram);
>> +	} else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
>>   		trace_cxl_memory_module(cxlmd, type, &evt->mem_module);
>>   	else
>>   		trace_cxl_generic_event(cxlmd, type, uuid, &evt->generic);
>>   }
>> -EXPORT_SYMBOL_NS_GPL(cxl_event_trace_record, CXL);
>> +EXPORT_SYMBOL_NS_GPL(cxl_event_handle_record, CXL);
>>
> 
> Why all the churn with changing the names of functions?

After adding cxl_report_poison(), these functions not only traces events 
but also contains event handler.  So rename them.

--
Thanks,
Ruan.

> 
> Ira
> 
>>   
>> -static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>> -				     enum cxl_event_log_type type,
>> -				     struct cxl_event_record_raw *record)
>> +static void __cxl_event_handle_record(struct cxl_memdev *cxlmd,
>> +				      enum cxl_event_log_type type,
>> +				      struct cxl_event_record_raw *record)
>>   {
>>   	enum cxl_event_type ev_type = CXL_CPER_EVENT_GENERIC;
>>   	const uuid_t *uuid = &record->id;
>> @@ -867,7 +958,7 @@ static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>>   	else if (uuid_equal(uuid, &CXL_EVENT_MEM_MODULE_UUID))
>>   		ev_type = CXL_CPER_EVENT_MEM_MODULE;
>>   
>> -	cxl_event_trace_record(cxlmd, type, ev_type, uuid, &record->event);
>> +	cxl_event_handle_record(cxlmd, type, ev_type, uuid, &record->event);
>>   }
>>   
>>   static int cxl_clear_event_record(struct cxl_memdev_state *mds,
>> @@ -979,8 +1070,8 @@ static void cxl_mem_get_records_log(struct cxl_memdev_state *mds,
>>   			break;
>>   
>>   		for (i = 0; i < nr_rec; i++)
>> -			__cxl_event_trace_record(cxlmd, type,
>> -						 &payload->records[i]);
>> +			__cxl_event_handle_record(cxlmd, type,
>> +						  &payload->records[i]);
>>   
>>   		if (payload->flags & CXL_GET_EVENT_FLAG_OVERFLOW)
>>   			trace_cxl_overflow(cxlmd, type, payload);
>> diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
>> index 36cee9c30ceb..ba1347de5651 100644
>> --- a/drivers/cxl/cxlmem.h
>> +++ b/drivers/cxl/cxlmem.h
>> @@ -822,10 +822,10 @@ void set_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>>   void clear_exclusive_cxl_commands(struct cxl_memdev_state *mds,
>>   				  unsigned long *cmds);
>>   void cxl_mem_get_event_records(struct cxl_memdev_state *mds, u32 status);
>> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>> -			    enum cxl_event_log_type type,
>> -			    enum cxl_event_type event_type,
>> -			    const uuid_t *uuid, union cxl_event *evt);
>> +void cxl_event_handle_record(struct cxl_memdev *cxlmd,
>> +			     enum cxl_event_log_type type,
>> +			     enum cxl_event_type event_type,
>> +			     const uuid_t *uuid, union cxl_event *evt);
>>   int cxl_set_timestamp(struct cxl_memdev_state *mds);
>>   int cxl_poison_state_init(struct cxl_memdev_state *mds);
>>   int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
>> diff --git a/include/linux/cxl-event.h b/include/linux/cxl-event.h
>> index 03fa6d50d46f..8189bed76c12 100644
>> --- a/include/linux/cxl-event.h
>> +++ b/include/linux/cxl-event.h
>> @@ -23,6 +23,20 @@ struct cxl_event_generic {
>>   	u8 data[CXL_EVENT_RECORD_DATA_LENGTH];
>>   } __packed;
>>   
>> +/*
>> + * Event transaction type
>> + * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
>> + */
>> +enum cxl_event_transaction_type {
>> +	CXL_EVENT_TRANSACTION_UNKNOWN = 0X00,
>> +	CXL_EVENT_TRANSACTION_READ,
>> +	CXL_EVENT_TRANSACTION_WRITE,
>> +	CXL_EVENT_TRANSACTION_SCAN_MEDIA,
>> +	CXL_EVENT_TRANSACTION_INJECT_POISON,
>> +	CXL_EVENT_TRANSACTION_MEDIA_SCRUB,
>> +	CXL_EVENT_TRANSACTION_MEDIA_MANAGEMENT,
>> +};
>> +
>>   /*
>>    * General Media Event Record
>>    * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
>> @@ -33,7 +47,7 @@ struct cxl_event_gen_media {
>>   	__le64 phys_addr;
>>   	u8 descriptor;
>>   	u8 type;
>> -	u8 transaction_type;
>> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>>   	u8 validity_flags[2];
>>   	u8 channel;
>>   	u8 rank;
>> @@ -52,7 +66,7 @@ struct cxl_event_dram {
>>   	__le64 phys_addr;
>>   	u8 descriptor;
>>   	u8 type;
>> -	u8 transaction_type;
>> +	u8 transaction_type;	/* enum cxl_event_transaction_type */
>>   	u8 validity_flags[2];
>>   	u8 channel;
>>   	u8 rank;
>> -- 
>> 2.34.1
>>
> 
>
Shiyang Ruan May 3, 2024, 11:32 a.m. UTC | #7
在 2024/4/24 2:40, Dan Williams 写道:
> Shiyang Ruan wrote:
>> Currently driver only traces cxl events, poison creation (for both vmem
>> and pmem type) on cxl memdev is silent.
> 
> As it should be.
> 
>> OS needs to be notified then it could handle poison pages in time.
> 
> No, it was always the case that latent poison is an "action optional"
> event. I am not understanding the justification for this approach. What
> breaks if the kernel does not forward events to memory_failure_queue()?

I think for type3(pmem) device, it should be handled like NVDIMM.  If 
there are processes or filesystems running on it, they could be notified 
then operate a friendly shutdown if POISON happens.

> 
> Consider that in the CPU consumption case that the firmware first path
> will do its own memory_failure_queue() and in the native case the MCE
> handler will take care of this. So that leaves pages that are accessed
> by DMA or background operation that encounter poison. Those are "action
> optional" scenarios and it is not clear to me how the driver tells the
> difference.

So for real CXL device, it always use FW-First path to notify such 
failure event?  Then, there is nothing to do with OS-First path?

> 
> This needs more precision on which agent is repsonsible for what level
> of reporting. The distribution of responsibility between ACPI GHES,
> EDAC, and the CXL driver is messy and I expect this changelog to
> demonstrate it understands all those considerations.

Ok, I'll try to understand them.

> 
>> Per CXL spec, the device error event could be signaled through
>> FW-First and OS-First methods.
>>
>> So, add poison creation event handler in OS-First method:
>>    - Qemu:
> 
> Why is QEMU relevant for this patch? QEMU is only a development vehicle
> the upstream enabling should be reference shipping or expected to be
> shipping hardware implementations.

Yes, but currently we don't have a real CXL device so developing and 
verification could only be done on Qemu with simulated CXL device.

> 
>>      - CXL device reports POISON creation event to OS by MSI by sending
>>        GMER/DER after injecting a poison record;
> 
> When you say "inject" here do you mean "add to the poison list if
> present". Because "inject" to me means the "Inject Poison" Memory Device
> Command.

It's a Qemu qmp command called "cxl-inject-poison", which only adds a 
given address,length record to CXL's poison list, doesn't send 
INJECT_POISON mbox command.

> 
>>    - CXL driver:
>>      a. parse the POISON event from GMER/DER;
>>      b. translate poisoned DPA to HPA (PFN);
>>      c. enqueue poisoned PFN to memory_failure's work queue;
>>
>> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
>> ---
>>   drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>>   drivers/cxl/cxlmem.h      |   8 +--
>>   include/linux/cxl-event.h |  18 +++++-
>>   3 files changed, 125 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
>> index f0f54aeccc87..76af0d73859d 100644
>> --- a/drivers/cxl/core/mbox.c
>> +++ b/drivers/cxl/core/mbox.c
>> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
>>   }
>>   EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>>   
>> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>> -			    enum cxl_event_log_type type,
>> -			    enum cxl_event_type event_type,
>> -			    const uuid_t *uuid, union cxl_event *evt)
>> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
>> +			      u64 dpa)
>>   {
>> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
>> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>> +	unsigned long pfn = PHYS_PFN(hpa);
>> +
>> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
>> +		return;
> 
> No need for this check, memory_failure_queue() is already stubbed out in
> the CONFIG_MEMORY_FAILURE=n case.
Yes, I'm overthinking it.

> 
>> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);
> 
> My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
> reported errors since action is only required for direct consumption
> events and those need not be reported through the device event queue.
Got it.

> 
> It would be useful to collaborate with a BIOS firmware engineer so that
> the kernel ends up with similar logic as is used to set CPER record
> severity, or at least understands why it would want to be different.
> See how ghes_handle_memory_failure() determines the
> memory_failure_queue() flags.

Ok, thanks for the advice.


--
Ruan.
Jonathan Cameron May 8, 2024, 4:15 p.m. UTC | #8
On Fri, 3 May 2024 18:42:31 +0800
Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:

> 在 2024/4/24 1:57, Ira Weiny 写道:
> > Shiyang Ruan wrote:  
> >> Currently driver only traces cxl events, poison creation (for both vmem
> >> and pmem type) on cxl memdev is silent.  OS needs to be notified then it
> >> could handle poison pages in time.  Per CXL spec, the device error event
> >> could be signaled through FW-First and OS-First methods.
> >>
> >> So, add poison creation event handler in OS-First method:
> >>    - Qemu:
> >>      - CXL device reports POISON creation event to OS by MSI by sending
> >>        GMER/DER after injecting a poison record;
> >>    - CXL driver:
> >>      a. parse the POISON event from GMER/DER;
> >>      b. translate poisoned DPA to HPA (PFN);
> >>      c. enqueue poisoned PFN to memory_failure's work queue;  
> > 
> > I'm a bit confused by the need for this patch.  Perhaps a bit more detail
> > here?  
> 
> Yes, I should have wrote more details.
> 
> I want to check and make sure the HWPOISON on a CXL device (type3) is 
> working properly.  For example, a FSDAX filesystem created on a 
> type3-pmem device, then it gets a POISON bit, the OS should be able to 
> handle this POISON event: find the relevant process
> 
> Currently I'm using Qemu with several simulated CXL devices, and using 
> poison injection API of Qemu to create POISON records, but OS isn't 
> notified.  Only when we actively call list POISON records (cxl list -L) 
> will the driver fetch them and log into trace events, then we see the 
> POISON records.  Memory failure wasn't triggered too.
> 

Indeed - QEMU emulation of this is not complete.  It should also be generating
the events. Ideally we'd even handle injecting silent poison (not yet detected
by the device) and have that generate the synchronous memory faults on an access.

> That's why I said "poison creation on cxl memdev is silent".  Per spec, 
> POISON creation should be notified to OS.  Since not familiar with 
> firmware part, I'm try adding this notification for OS-First.
> 
> > 
> > More comments below.
> >   
> >>
> >> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> >> ---
> >>   drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
> >>   drivers/cxl/cxlmem.h      |   8 +--
> >>   include/linux/cxl-event.h |  18 +++++-
> >>   3 files changed, 125 insertions(+), 20 deletions(-)
> >>
> >> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> >> index f0f54aeccc87..76af0d73859d 100644
> >> --- a/drivers/cxl/core/mbox.c
> >> +++ b/drivers/cxl/core/mbox.c
> >> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
> >>   }
> >>   EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
> >>   
> >> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
> >> -			    enum cxl_event_log_type type,
> >> -			    enum cxl_event_type event_type,
> >> -			    const uuid_t *uuid, union cxl_event *evt)
> >> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
> >> +			      u64 dpa)
> >>   {
> >> -	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
> >> +	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
> >> +	unsigned long pfn = PHYS_PFN(hpa);
> >> +
> >> +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
> >> +		return;
> >> +
> >> +	memory_failure_queue(pfn, MF_ACTION_REQUIRED);  
> > 
> > I thought that ras daemon was supposed to take care of this when the trace
> > event occurred.  Alison is working on the HPA data for that path.  
> 
> It seems to save CXL trace events/memory-failures to DB and report to 
> others, but cannot let OS call memory failure.

Interesting question of whose problem this is.  For corrected
errors it's policy stuff that belongs in userspace, but for known
memory failure, it might want to be in kernel.

Shiju (+Cc) pointed me a the existing rasdaemon handling for corrected
errors (statistics get too bad, so memory offlined) 
https://github.com/mchehab/rasdaemon/commit/9ae6b70effb8adc9572debc800b8e16173f74bb8

Poison detection via scrub though is reported via CPER Memory Error Section with
"Scrub uncorrected error" set. That triggers apei handling. On
x86 looks to be apei_mce_report_mem_error(). Also triggers
the ghes_do_memory_failure_path() and ultimately memory_failure().

Conveniently there was a patch fixing the sync path last year
that includes info on what happens in async case.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a70297d2213253853e95f5b49651f924990c6d3b

"In addition, for aysnchronous errors, kernel will notify the
process who owns the poisoned page by sending SIGBUS with BUS_MCERRR_A0
in early kill mode."

So I think the kernel should probably do the same for CXL poison
error records when it gets them.

Jonathan
Shiyang Ruan May 21, 2024, 5:35 a.m. UTC | #9
在 2024/5/3 19:32, Shiyang Ruan 写道:
> 
> 
> 在 2024/4/24 2:40, Dan Williams 写道:
>> Shiyang Ruan wrote:
>>> Currently driver only traces cxl events, poison creation (for both vmem
>>> and pmem type) on cxl memdev is silent.
>>
>> As it should be.
>>
>>> OS needs to be notified then it could handle poison pages in time.
>>
>> No, it was always the case that latent poison is an "action optional"
>> event. I am not understanding the justification for this approach. What
>> breaks if the kernel does not forward events to memory_failure_queue()?
> 
> I think for type3(pmem) device, it should be handled like NVDIMM.  If 
> there are processes or filesystems running on it, they could be notified 
> then operate a friendly shutdown if POISON happens.
> 
>>
>> Consider that in the CPU consumption case that the firmware first path
>> will do its own memory_failure_queue() and in the native case the MCE
>> handler will take care of this. So that leaves pages that are accessed
>> by DMA or background operation that encounter poison. Those are "action
>> optional" scenarios and it is not clear to me how the driver tells the
>> difference.
> 
> So for real CXL device, it always use FW-First path to notify such 
> failure event?  Then, there is nothing to do with OS-First path?
> 
>>
>> This needs more precision on which agent is repsonsible for what level
>> of reporting. The distribution of responsibility between ACPI GHES,
>> EDAC, and the CXL driver is messy and I expect this changelog to
>> demonstrate it understands all those considerations.
> 
> Ok, I'll try to understand them.

Hi Dan,

I checked the GHES, EDAC codes. I think they belong to FW-First path. 
GHES polls mem errors, then
  1. report by EDAC
  2. construct a MCE, mce_log(), and handle in work queue
  3. queue it into memory_failure right now if needed (sync, ...)
And community is adding CXL FW-First trace support[1].

But in OS-First path, error record is sent to CXL driver via MSI, it 
won't conflict with FW-First path, I think.

[1] 
https://lore.kernel.org/linux-cxl/43ab39e9-c9c2-bfe4-7d1c-bad462221c9c@amd.com/T/#t

> 
>>
>>> Per CXL spec, the device error event could be signaled through
>>> FW-First and OS-First methods.
>>>
>>> So, add poison creation event handler in OS-First method:
>>>    - Qemu:
>>
>> Why is QEMU relevant for this patch? QEMU is only a development vehicle
>> the upstream enabling should be reference shipping or expected to be
>> shipping hardware implementations.
> 
> Yes, but currently we don't have a real CXL device so developing and 
> verification could only be done on Qemu with simulated CXL device.
> 
>>
>>>      - CXL device reports POISON creation event to OS by MSI by sending
>>>        GMER/DER after injecting a poison record;
>>
>> When you say "inject" here do you mean "add to the poison list if
>> present". Because "inject" to me means the "Inject Poison" Memory Device
>> Command.
> 
> It's a Qemu qmp command called "cxl-inject-poison", which only adds a 
> given address,length record to CXL's poison list, doesn't send 
> INJECT_POISON mbox command.
> 
>>
>>>    - CXL driver:
>>>      a. parse the POISON event from GMER/DER;
>>>      b. translate poisoned DPA to HPA (PFN);
>>>      c. enqueue poisoned PFN to memory_failure's work queue;
>>>
>>> Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
>>> ---
>>>   drivers/cxl/core/mbox.c   | 119 +++++++++++++++++++++++++++++++++-----
>>>   drivers/cxl/cxlmem.h      |   8 +--
>>>   include/linux/cxl-event.h |  18 +++++-
>>>   3 files changed, 125 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
>>> index f0f54aeccc87..76af0d73859d 100644
>>> --- a/drivers/cxl/core/mbox.c
>>> +++ b/drivers/cxl/core/mbox.c
>>> @@ -837,25 +837,116 @@ int cxl_enumerate_cmds(struct cxl_memdev_state 
>>> *mds)
>>>   }
>>>   EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
>>> -void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
>>> -                enum cxl_event_log_type type,
>>> -                enum cxl_event_type event_type,
>>> -                const uuid_t *uuid, union cxl_event *evt)
>>> +static void cxl_report_poison(struct cxl_memdev *cxlmd, struct 
>>> cxl_region *cxlr,
>>> +                  u64 dpa)
>>>   {
>>> -    if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
>>> +    u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
>>> +    unsigned long pfn = PHYS_PFN(hpa);
>>> +
>>> +    if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
>>> +        return;
>>
>> No need for this check, memory_failure_queue() is already stubbed out in
>> the CONFIG_MEMORY_FAILURE=n case.
> Yes, I'm overthinking it.
> 
>>
>>> +    memory_failure_queue(pfn, MF_ACTION_REQUIRED);
>>
>> My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
>> reported errors since action is only required for direct consumption
>> events and those need not be reported through the device event queue.
> Got it.

I'm not very sure about 'Host write/read' type.  In my opinion, these 
two types of event should be sent from device when CPU is accessing a 
bad memory address, they could be thought of a sync event which needs 
the 'MF_ACTION_REQUIRED' flag.  Then, we can determine the flag by the 
types like this:
- CXL_EVENT_TRANSACTION_READ | CXL_EVENT_TRANSACTION_WRITE
                                               => MF_ACTION_REQUIRED
- CXL_EVENT_TRANSACTION_INJECT_POISON         => MF_SW_SIMULATED
- others                                      => 0


--
Thanks,
Ruan.

> 
>>
>> It would be useful to collaborate with a BIOS firmware engineer so that
>> the kernel ends up with similar logic as is used to set CPER record
>> severity, or at least understands why it would want to be different.
>> See how ghes_handle_memory_failure() determines the
>> memory_failure_queue() flags.
> 
> Ok, thanks for the advice.
> 
> 
> -- 
> Ruan.
> 
> 
>
Dan Williams May 22, 2024, 6:45 a.m. UTC | #10
Shiyang Ruan wrote:
[..]
> >> My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
> >> reported errors since action is only required for direct consumption
> >> events and those need not be reported through the device event queue.
> > Got it.
> 
> I'm not very sure about 'Host write/read' type.  In my opinion, these 
> two types of event should be sent from device when CPU is accessing a 
> bad memory address, they could be thought of a sync event which needs 

Hmm, no that's not my understanding of a sync event. I expect when error
notifications are synchronous the CPU is guaranteed not to make forward
progress past the point of encountering the error. MSI-signaled
component-events are always asynchronous by that definition because the
CPU is free running while the interrupt is in-flight.

> the 'MF_ACTION_REQUIRED' flag.  Then, we can determine the flag by the 
> types like this:
> - CXL_EVENT_TRANSACTION_READ | CXL_EVENT_TRANSACTION_WRITE
>                                                => MF_ACTION_REQUIRED
> - CXL_EVENT_TRANSACTION_INJECT_POISON         => MF_SW_SIMULATED
> - others                                      => 0

I doubt any reasonable policy can be inferred from the transaction type.
Consider that the CPU itself does not take a sychronous exception when
writes encounter poison. At most those are flagged via CMCI
(corrected machine check interrupt). The only events that cause
exceptions are CPU reads that consume poison. The device has no idea
whether read events are coming from a CPU or a DMA event.

MF_SW_SIMULATED is purely for software simulated poison events as
injected poison can stil cause system fatal damage if the poison is
ingested in an unrecoverable path.

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.
Shiyang Ruan May 24, 2024, 3:15 p.m. UTC | #11
在 2024/5/22 14:45, Dan Williams 写道:
> Shiyang Ruan wrote:
> [..]
>>>> My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
>>>> reported errors since action is only required for direct consumption
>>>> events and those need not be reported through the device event queue.
>>> Got it.
>>
>> I'm not very sure about 'Host write/read' type.  In my opinion, these
>> two types of event should be sent from device when CPU is accessing a
>> bad memory address, they could be thought of a sync event which needs
> 
> Hmm, no that's not my understanding of a sync event. I expect when error
> notifications are synchronous the CPU is guaranteed not to make forward
> progress past the point of encountering the error. MSI-signaled
> component-events are always asynchronous by that definition because the
> CPU is free running while the interrupt is in-flight.

Understood.  In OS-First path, it couldn't be a sync event.

> 
>> the 'MF_ACTION_REQUIRED' flag.  Then, we can determine the flag by the
>> types like this:
>> - CXL_EVENT_TRANSACTION_READ | CXL_EVENT_TRANSACTION_WRITE
>>                                                 => MF_ACTION_REQUIRED
>> - CXL_EVENT_TRANSACTION_INJECT_POISON         => MF_SW_SIMULATED
>> - others                                      => 0
> 
> I doubt any reasonable policy can be inferred from the transaction type.
> Consider that the CPU itself does not take a sychronous exception when
> writes encounter poison. At most those are flagged via CMCI
> (corrected machine check interrupt). The only events that cause
> exceptions are CPU reads that consume poison. The device has no idea
> whether read events are coming from a CPU or a DMA event.
> 
> MF_SW_SIMULATED is purely for software simulated poison events as
> injected poison can stil cause system fatal damage if the poison is
> ingested in an unrecoverable path.
> 
> 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.

I think my previous understanding about MCE was wrong.  Here is my 
current understanding after some research:

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 
-> MCE.  This is the second one.  These two MCEs represent the same 
poisoning page, which is the so-called "duplicate report", right?  Now, 
the memory_failure() handling I'm trying to add in OS-First path, is 
also another duplicate report.

So, the primary issue to be solved is the second MCE report.  As you 
suggested, make it a optional action.

Please correct me if I'm wrong.  Thank you very much!

--
Ruan.
Shiyang Ruan May 28, 2024, 10:13 a.m. UTC | #12
在 2024/5/24 23:15, Shiyang Ruan 写道:
> 
> 
> 在 2024/5/22 14:45, Dan Williams 写道:
>> Shiyang Ruan wrote:
>> [..]
>>>>> My expectation is MF_ACTION_REQUIRED is not appropriate for CXL event
>>>>> reported errors since action is only required for direct consumption
>>>>> events and those need not be reported through the device event queue.
>>>> Got it.
>>>
>>> I'm not very sure about 'Host write/read' type.  In my opinion, these
>>> two types of event should be sent from device when CPU is accessing a
>>> bad memory address, they could be thought of a sync event which needs
>>
>> Hmm, no that's not my understanding of a sync event. I expect when error
>> notifications are synchronous the CPU is guaranteed not to make forward
>> progress past the point of encountering the error. MSI-signaled
>> component-events are always asynchronous by that definition because the
>> CPU is free running while the interrupt is in-flight.
> 
> Understood.  In OS-First path, it couldn't be a sync event.
> 
>>
>>> the 'MF_ACTION_REQUIRED' flag.  Then, we can determine the flag by the
>>> types like this:
>>> - CXL_EVENT_TRANSACTION_READ | CXL_EVENT_TRANSACTION_WRITE
>>>                                                 => MF_ACTION_REQUIRED
>>> - CXL_EVENT_TRANSACTION_INJECT_POISON         => MF_SW_SIMULATED
>>> - others                                      => 0
>>
>> I doubt any reasonable policy can be inferred from the transaction type.
>> Consider that the CPU itself does not take a sychronous exception when
>> writes encounter poison. At most those are flagged via CMCI
>> (corrected machine check interrupt). The only events that cause
>> exceptions are CPU reads that consume poison. The device has no idea
>> whether read events are coming from a CPU or a DMA event.
>>
>> MF_SW_SIMULATED is purely for software simulated poison events as
>> injected poison can stil cause system fatal damage if the poison is
>> ingested in an unrecoverable path.
>>
>> 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.
> 
> I think my previous understanding about MCE was wrong.  Here is my 
> current understanding after some research:
> 
> 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 
> -> MCE.  This is the second one.  These two MCEs represent the same 
> poisoning page, which is the so-called "duplicate report", right?  Now, 
> the memory_failure() handling I'm trying to add in OS-First path, is 
> also another duplicate report.

Let me summarize:
CPU accesses bad memory on CXL device, then
-> MCE (INT18), *always* report (1)
-> * FW-First (implemented)
      -> CXL device -> FW -> OS:ACPI->APEI->GHES -> MCE (2)
    * OS-First (not implemented yet, I'm working on it)
      -> CXL device -> MSI -> OS:CXL driver -> memory_failure() (2)
Then (1) and (2) are duplicated.

Looking forward your comment.  Thanks in advance!


--
Ruan.

> 
> So, the primary issue to be solved is the second MCE report.  As you 
> suggested, make it a optional action.
> 
> Please correct me if I'm wrong.  Thank you very much!
> 
> -- 
> Ruan.
diff mbox series

Patch

diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
index f0f54aeccc87..76af0d73859d 100644
--- a/drivers/cxl/core/mbox.c
+++ b/drivers/cxl/core/mbox.c
@@ -837,25 +837,116 @@  int cxl_enumerate_cmds(struct cxl_memdev_state *mds)
 }
 EXPORT_SYMBOL_NS_GPL(cxl_enumerate_cmds, CXL);
 
-void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
-			    enum cxl_event_log_type type,
-			    enum cxl_event_type event_type,
-			    const uuid_t *uuid, union cxl_event *evt)
+static void cxl_report_poison(struct cxl_memdev *cxlmd, struct cxl_region *cxlr,
+			      u64 dpa)
 {
-	if (event_type == CXL_CPER_EVENT_GEN_MEDIA)
+	u64 hpa = cxl_trace_hpa(cxlr, cxlmd, dpa);
+	unsigned long pfn = PHYS_PFN(hpa);
+
+	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
+		return;
+
+	memory_failure_queue(pfn, MF_ACTION_REQUIRED);
+}
+
+static int __cxl_report_poison(struct device *dev, void *arg)
+{
+	struct cxl_endpoint_decoder *cxled;
+	struct cxl_memdev *cxlmd;
+	u64 dpa = *(u64 *)arg;
+
+	cxled = to_cxl_endpoint_decoder(dev);
+	if (!cxled || !cxled->dpa_res || !resource_size(cxled->dpa_res))
+		return 0;
+
+	if (cxled->mode == CXL_DECODER_MIXED) {
+		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
+		return 0;
+	}
+
+	if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
+		return 0;
+
+	cxlmd = cxled_to_memdev(cxled);
+	cxl_report_poison(cxlmd, cxled->cxld.region, dpa);
+
+	return 1;
+}
+
+static void cxl_event_handle_poison(struct cxl_memdev *cxlmd, u64 dpa)
+{
+	struct cxl_port *port = cxlmd->endpoint;
+
+	/*
+	 * No region is mapped to this endpoint, that is to say no HPA is
+	 * mapped.
+	 */
+	if (!port || !is_cxl_endpoint(port) ||
+	    cxl_num_decoders_committed(port) == 0)
+		return;
+
+	device_for_each_child(&port->dev, &dpa, __cxl_report_poison);
+}
+
+static void cxl_event_handle_general_media(struct cxl_memdev *cxlmd,
+					   enum cxl_event_log_type type,
+					   struct cxl_event_gen_media *rec)
+{
+	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
+
+	if (type == CXL_EVENT_TYPE_FAIL) {
+		switch (rec->transaction_type) {
+		case CXL_EVENT_TRANSACTION_READ:
+		case CXL_EVENT_TRANSACTION_WRITE:
+		case CXL_EVENT_TRANSACTION_INJECT_POISON:
+			cxl_event_handle_poison(cxlmd, dpa);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
+static void cxl_event_handle_dram(struct cxl_memdev *cxlmd,
+				  enum cxl_event_log_type type,
+				  struct cxl_event_dram *rec)
+{
+	u64 dpa = le64_to_cpu(rec->phys_addr) & CXL_DPA_MASK;
+
+	if (type == CXL_EVENT_TYPE_FAIL) {
+		switch (rec->transaction_type) {
+		case CXL_EVENT_TRANSACTION_READ:
+		case CXL_EVENT_TRANSACTION_WRITE:
+		case CXL_EVENT_TRANSACTION_INJECT_POISON:
+			cxl_event_handle_poison(cxlmd, dpa);
+			break;
+		default:
+			break;
+		}
+	}
+}
+
+void cxl_event_handle_record(struct cxl_memdev *cxlmd,
+			     enum cxl_event_log_type type,
+			     enum cxl_event_type event_type,
+			     const uuid_t *uuid, union cxl_event *evt)
+{
+	if (event_type == CXL_CPER_EVENT_GEN_MEDIA) {
 		trace_cxl_general_media(cxlmd, type, &evt->gen_media);
-	else if (event_type == CXL_CPER_EVENT_DRAM)
+		cxl_event_handle_general_media(cxlmd, type, &evt->gen_media);
+	} else if (event_type == CXL_CPER_EVENT_DRAM) {
 		trace_cxl_dram(cxlmd, type, &evt->dram);
-	else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
+		cxl_event_handle_dram(cxlmd, type, &evt->dram);
+	} else if (event_type == CXL_CPER_EVENT_MEM_MODULE)
 		trace_cxl_memory_module(cxlmd, type, &evt->mem_module);
 	else
 		trace_cxl_generic_event(cxlmd, type, uuid, &evt->generic);
 }
-EXPORT_SYMBOL_NS_GPL(cxl_event_trace_record, CXL);
+EXPORT_SYMBOL_NS_GPL(cxl_event_handle_record, CXL);
 
-static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
-				     enum cxl_event_log_type type,
-				     struct cxl_event_record_raw *record)
+static void __cxl_event_handle_record(struct cxl_memdev *cxlmd,
+				      enum cxl_event_log_type type,
+				      struct cxl_event_record_raw *record)
 {
 	enum cxl_event_type ev_type = CXL_CPER_EVENT_GENERIC;
 	const uuid_t *uuid = &record->id;
@@ -867,7 +958,7 @@  static void __cxl_event_trace_record(const struct cxl_memdev *cxlmd,
 	else if (uuid_equal(uuid, &CXL_EVENT_MEM_MODULE_UUID))
 		ev_type = CXL_CPER_EVENT_MEM_MODULE;
 
-	cxl_event_trace_record(cxlmd, type, ev_type, uuid, &record->event);
+	cxl_event_handle_record(cxlmd, type, ev_type, uuid, &record->event);
 }
 
 static int cxl_clear_event_record(struct cxl_memdev_state *mds,
@@ -979,8 +1070,8 @@  static void cxl_mem_get_records_log(struct cxl_memdev_state *mds,
 			break;
 
 		for (i = 0; i < nr_rec; i++)
-			__cxl_event_trace_record(cxlmd, type,
-						 &payload->records[i]);
+			__cxl_event_handle_record(cxlmd, type,
+						  &payload->records[i]);
 
 		if (payload->flags & CXL_GET_EVENT_FLAG_OVERFLOW)
 			trace_cxl_overflow(cxlmd, type, payload);
diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
index 36cee9c30ceb..ba1347de5651 100644
--- a/drivers/cxl/cxlmem.h
+++ b/drivers/cxl/cxlmem.h
@@ -822,10 +822,10 @@  void set_exclusive_cxl_commands(struct cxl_memdev_state *mds,
 void clear_exclusive_cxl_commands(struct cxl_memdev_state *mds,
 				  unsigned long *cmds);
 void cxl_mem_get_event_records(struct cxl_memdev_state *mds, u32 status);
-void cxl_event_trace_record(const struct cxl_memdev *cxlmd,
-			    enum cxl_event_log_type type,
-			    enum cxl_event_type event_type,
-			    const uuid_t *uuid, union cxl_event *evt);
+void cxl_event_handle_record(struct cxl_memdev *cxlmd,
+			     enum cxl_event_log_type type,
+			     enum cxl_event_type event_type,
+			     const uuid_t *uuid, union cxl_event *evt);
 int cxl_set_timestamp(struct cxl_memdev_state *mds);
 int cxl_poison_state_init(struct cxl_memdev_state *mds);
 int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
diff --git a/include/linux/cxl-event.h b/include/linux/cxl-event.h
index 03fa6d50d46f..8189bed76c12 100644
--- a/include/linux/cxl-event.h
+++ b/include/linux/cxl-event.h
@@ -23,6 +23,20 @@  struct cxl_event_generic {
 	u8 data[CXL_EVENT_RECORD_DATA_LENGTH];
 } __packed;
 
+/*
+ * Event transaction type
+ * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
+ */
+enum cxl_event_transaction_type {
+	CXL_EVENT_TRANSACTION_UNKNOWN = 0X00,
+	CXL_EVENT_TRANSACTION_READ,
+	CXL_EVENT_TRANSACTION_WRITE,
+	CXL_EVENT_TRANSACTION_SCAN_MEDIA,
+	CXL_EVENT_TRANSACTION_INJECT_POISON,
+	CXL_EVENT_TRANSACTION_MEDIA_SCRUB,
+	CXL_EVENT_TRANSACTION_MEDIA_MANAGEMENT,
+};
+
 /*
  * General Media Event Record
  * CXL rev 3.0 Section 8.2.9.2.1.1; Table 8-43
@@ -33,7 +47,7 @@  struct cxl_event_gen_media {
 	__le64 phys_addr;
 	u8 descriptor;
 	u8 type;
-	u8 transaction_type;
+	u8 transaction_type;	/* enum cxl_event_transaction_type */
 	u8 validity_flags[2];
 	u8 channel;
 	u8 rank;
@@ -52,7 +66,7 @@  struct cxl_event_dram {
 	__le64 phys_addr;
 	u8 descriptor;
 	u8 type;
-	u8 transaction_type;
+	u8 transaction_type;	/* enum cxl_event_transaction_type */
 	u8 validity_flags[2];
 	u8 channel;
 	u8 rank;