[PATCH v7 0/6] CXL Poison List Retrieval & Tracing

[PATCH v7 0/6] CXL Poison List Retrieval & Tracing

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

This is not targeting the 6.3 merge window.

Changes in v7:
- Update commit log wrt poison read for unmapped after mapped (Jonathan)
- Update comment wrt decoder modes and pmem/ram separation (Jonathan)
- Add dev_dbg() if mixed mode decoder appears
- Read poison on skip range first (Jonathan)
- trace: order fields (memdev, host, serial)
- trace: simplify host name get
- trace: use typical tabs
Link to v6: https://lore.kernel.org/linux-cxl/cover.1675983077.git.alison.schofield@intel.com/

Add support for retrieving device poison lists and store the returned
error records as kernel trace events.

The handling of the poison list is guided by the CXL 3.0 Specification
Section 8.2.9.8.4.1. [1] 

Example:
$ echo 1 > /sys/bus/cxl/devices/mem0/trigger_poison_list
cxl_poison: memdev=mem0 host=cxl_mem.0 serial=0 region=region4 region_uuid=117b2cf4-b160-4090-9361-ba31b9649317 hpa=0xf0d0000000 dpa=0x40000000 length=0x40 source=Internal flags= overflow_time=0

[1]: https://www.computeexpresslink.org/download-the-specification


Alison Schofield (6):
  cxl/mbox: Add GET_POISON_LIST mailbox command
  cxl/trace: Add TRACE support for CXL media-error records
  cxl/memdev: Add trigger_poison_list sysfs attribute
  cxl/region: Provide region info to the cxl_poison trace event
  cxl/trace: Add an HPA to cxl_poison trace events
  tools/testing/cxl: Mock support for Get Poison List

 Documentation/ABI/testing/sysfs-bus-cxl | 14 ++++
 drivers/cxl/core/core.h                 |  5 ++
 drivers/cxl/core/mbox.c                 | 74 +++++++++++++++++++
 drivers/cxl/core/memdev.c               | 68 ++++++++++++++++++
 drivers/cxl/core/region.c               | 89 +++++++++++++++++++++++
 drivers/cxl/core/trace.c                | 94 +++++++++++++++++++++++++
 drivers/cxl/core/trace.h                | 91 ++++++++++++++++++++++++
 drivers/cxl/cxlmem.h                    | 69 +++++++++++++++++-
 drivers/cxl/pci.c                       |  4 ++
 tools/testing/cxl/test/mem.c            | 42 +++++++++++
 10 files changed, 549 insertions(+), 1 deletion(-)


base-commit: a5fcd228ca1db9810ba1ed461c90b6ee933b9daf
-- 
2.37.3

[PATCH v7 1/6] cxl/mbox: Add GET_POISON_LIST mailbox command

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

CXL devices maintain a list of locations that are poisoned or result
in poison if the addresses are accessed by the host.

Per the spec (CXL 3.0 8.2.9.8.4.1), the device returns this Poison
list as a set of  Media Error Records that include the source of the
error, the starting device physical address and length. The length is
the number of adjacent DPAs in the record and is in units of 64 bytes.

Retrieve the poison list.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 drivers/cxl/core/mbox.c | 72 +++++++++++++++++++++++++++++++++++++++++
 drivers/cxl/cxlmem.h    | 67 ++++++++++++++++++++++++++++++++++++++
 drivers/cxl/pci.c       |  4 +++
 3 files changed, 143 insertions(+)

diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
index fc7631bb1c24..faa61e5f541f 100644
--- a/drivers/cxl/core/mbox.c
+++ b/drivers/cxl/core/mbox.c
@@ -5,6 +5,8 @@
 #include <linux/debugfs.h>
 #include <linux/ktime.h>
 #include <linux/mutex.h>
+#include <asm/unaligned.h>
+#include <cxlpci.h>
 #include <cxlmem.h>
 #include <cxl.h>
 
@@ -994,6 +996,7 @@ int cxl_dev_state_identify(struct cxl_dev_state *cxlds)
 	/* See CXL 2.0 Table 175 Identify Memory Device Output Payload */
 	struct cxl_mbox_identify id;
 	struct cxl_mbox_cmd mbox_cmd;
+	u32 val;
 	int rc;
 
 	mbox_cmd = (struct cxl_mbox_cmd) {
@@ -1017,6 +1020,11 @@ int cxl_dev_state_identify(struct cxl_dev_state *cxlds)
 	cxlds->lsa_size = le32_to_cpu(id.lsa_size);
 	memcpy(cxlds->firmware_version, id.fw_revision, sizeof(id.fw_revision));
 
+	if (test_bit(CXL_MEM_COMMAND_ID_GET_POISON, cxlds->enabled_cmds)) {
+		val = get_unaligned_le24(id.poison_list_max_mer);
+		cxlds->poison.max_errors = min_t(u32, val, CXL_POISON_LIST_MAX);
+	}
+
 	return 0;
 }
 EXPORT_SYMBOL_NS_GPL(cxl_dev_state_identify, CXL);
@@ -1107,6 +1115,70 @@ int cxl_set_timestamp(struct cxl_dev_state *cxlds)
 }
 EXPORT_SYMBOL_NS_GPL(cxl_set_timestamp, CXL);
 
+int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
+		       struct cxl_region *cxlr)
+{
+	struct cxl_dev_state *cxlds = cxlmd->cxlds;
+	struct cxl_mbox_poison_payload_out *po;
+	struct cxl_mbox_poison_payload_in pi;
+	struct cxl_mbox_cmd mbox_cmd;
+	int nr_records = 0;
+	int rc;
+
+	rc = mutex_lock_interruptible(&cxlds->poison.lock);
+	if (rc)
+		return rc;
+
+	po = cxlds->poison.payload_out;
+	pi.offset = cpu_to_le64(offset);
+	pi.length = cpu_to_le64(len / CXL_POISON_LEN_MULT);
+
+	mbox_cmd = (struct cxl_mbox_cmd) {
+		.opcode = CXL_MBOX_OP_GET_POISON,
+		.size_in = sizeof(pi),
+		.payload_in = &pi,
+		.size_out = cxlds->payload_size,
+		.payload_out = po,
+		.min_out = struct_size(po, record, 0),
+	};
+
+	do {
+		rc = cxl_internal_send_cmd(cxlds, &mbox_cmd);
+		if (rc)
+			break;
+
+		/* TODO TRACE the media error records */
+
+		/* Protect against an uncleared _FLAG_MORE */
+		nr_records = nr_records + le16_to_cpu(po->count);
+		if (nr_records >= cxlds->poison.max_errors) {
+			dev_dbg(&cxlmd->dev, "Max Error Records reached: %d\n",
+				nr_records);
+			break;
+		}
+	} while (po->flags & CXL_POISON_FLAG_MORE);
+
+	mutex_unlock(&cxlds->poison.lock);
+	return rc;
+}
+EXPORT_SYMBOL_NS_GPL(cxl_mem_get_poison, CXL);
+
+int cxl_poison_state_init(struct cxl_dev_state *cxlds)
+{
+	if (!test_bit(CXL_MEM_COMMAND_ID_GET_POISON, cxlds->enabled_cmds))
+		return 0;
+
+	cxlds->poison.payload_out = devm_kzalloc(cxlds->dev,
+						 cxlds->payload_size,
+						 GFP_KERNEL);
+	if (!cxlds->poison.payload_out)
+		return -ENOMEM;
+
+	mutex_init(&cxlds->poison.lock);
+	return 0;
+}
+EXPORT_SYMBOL_NS_GPL(cxl_poison_state_init, CXL);
+
 struct cxl_dev_state *cxl_dev_state_create(struct device *dev)
 {
 	struct cxl_dev_state *cxlds;
diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
index c6c560c67a8a..c0feb43b585d 100644
--- a/drivers/cxl/cxlmem.h
+++ b/drivers/cxl/cxlmem.h
@@ -215,6 +215,24 @@ struct cxl_event_state {
 	struct mutex log_lock;
 };
 
+/**
+ * struct cxl_poison_state - Poison list retrieval
+ *
+ * @max_errors: Maximum media error records held in device cache
+ * @payload_out: The poison list payload returned by device
+ * @lock: Protect reads of the poison list
+ *
+ * Reads of the poison list are synchronized to ensure that a reader
+ * does not get an incomplete list because their request overlapped
+ * (was interrupted or preceded by) another read request of the same
+ * DPA range. CXL Spec 3.0 Section 8.2.9.8.4.1
+ */
+struct cxl_poison_state {
+	u32 max_errors;
+	struct cxl_mbox_poison_payload_out *payload_out;
+	struct mutex lock;  /* Protect reads of poison list */
+};
+
 /**
  * struct cxl_dev_state - The driver device state
  *
@@ -250,6 +268,7 @@ struct cxl_event_state {
  * @info: Cached DVSEC information about the device.
  * @serial: PCIe Device Serial Number
  * @doe_mbs: PCI DOE mailbox array
+ * @poison: poison list retrieval info
  * @mbox_send: @dev specific transport for transmitting mailbox commands
  *
  * See section 8.2.9.5.2 Capacity Configuration and Label Storage for
@@ -289,6 +308,7 @@ struct cxl_dev_state {
 	struct xarray doe_mbs;
 
 	struct cxl_event_state event;
+	struct cxl_poison_state poison;
 
 	int (*mbox_send)(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *cmd);
 };
@@ -537,6 +557,50 @@ struct cxl_mbox_set_timestamp_in {
 
 } __packed;
 
+/* Get Poison List  CXL 3.0 Spec 8.2.9.8.4.1 */
+struct cxl_mbox_poison_payload_in {
+	__le64 offset;
+	__le64 length;
+} __packed;
+
+struct cxl_mbox_poison_payload_out {
+	u8 flags;
+	u8 rsvd1;
+	__le64 overflow_t;
+	__le16 count;
+	u8 rsvd2[20];
+	struct cxl_poison_record {
+		__le64 address;
+		__le32 length;
+		__le32 rsvd;
+	} __packed record[];
+} __packed;
+
+/*
+ * Get Poison List address field encodes the starting
+ * address of poison, and the source of the poison.
+ */
+#define CXL_POISON_START_MASK		GENMASK_ULL(63, 6)
+#define CXL_POISON_SOURCE_MASK		GENMASK(2, 0)
+
+/* Get Poison List record length is in units of 64 bytes */
+#define CXL_POISON_LEN_MULT	64
+
+/* Kernel defined maximum for a list of poison errors */
+#define CXL_POISON_LIST_MAX	1024
+
+/* Get Poison List: Payload out flags */
+#define CXL_POISON_FLAG_MORE            BIT(0)
+#define CXL_POISON_FLAG_OVERFLOW        BIT(1)
+#define CXL_POISON_FLAG_SCANNING        BIT(2)
+
+/* Get Poison List: Poison Source */
+#define CXL_POISON_SOURCE_UNKNOWN	0
+#define CXL_POISON_SOURCE_EXTERNAL	1
+#define CXL_POISON_SOURCE_INTERNAL	2
+#define CXL_POISON_SOURCE_INJECTED	3
+#define CXL_POISON_SOURCE_VENDOR	7
+
 /**
  * struct cxl_mem_command - Driver representation of a memory device command
  * @info: Command information as it exists for the UAPI
@@ -607,6 +671,9 @@ void set_exclusive_cxl_commands(struct cxl_dev_state *cxlds, unsigned long *cmds
 void clear_exclusive_cxl_commands(struct cxl_dev_state *cxlds, unsigned long *cmds);
 void cxl_mem_get_event_records(struct cxl_dev_state *cxlds, u32 status);
 int cxl_set_timestamp(struct cxl_dev_state *cxlds);
+int cxl_poison_state_init(struct cxl_dev_state *cxlds);
+int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
+		       struct cxl_region *cxlr);
 
 #ifdef CONFIG_CXL_SUSPEND
 void cxl_mem_active_inc(void);
diff --git a/drivers/cxl/pci.c b/drivers/cxl/pci.c
index 60b23624d167..827ea0895778 100644
--- a/drivers/cxl/pci.c
+++ b/drivers/cxl/pci.c
@@ -769,6 +769,10 @@ static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	if (rc)
 		return rc;
 
+	rc = cxl_poison_state_init(cxlds);
+	if (rc)
+		return rc;
+
 	rc = cxl_dev_state_identify(cxlds);
 	if (rc)
 		return rc;
-- 
2.37.3

[PATCH v7 2/6] cxl/trace: Add TRACE support for CXL media-error records

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

CXL devices may support the retrieval of a device poison list.
Add a new trace event that the CXL subsystem may use to log
the media-error records returned in the poison list.

Log each media-error record as a trace event of type 'cxl_poison'.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 drivers/cxl/core/mbox.c  |  4 +-
 drivers/cxl/core/trace.h | 84 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 87 insertions(+), 1 deletion(-)

diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
index faa61e5f541f..a3da8b4e92a4 100644
--- a/drivers/cxl/core/mbox.c
+++ b/drivers/cxl/core/mbox.c
@@ -1147,7 +1147,9 @@ int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
 		if (rc)
 			break;
 
-		/* TODO TRACE the media error records */
+		for (int i = 0; i < le16_to_cpu(po->count); i++)
+			trace_cxl_poison(cxlmd, cxlr, &po->record[i],
+					 po->flags, po->overflow_t);
 
 		/* Protect against an uncleared _FLAG_MORE */
 		nr_records = nr_records + le16_to_cpu(po->count);
diff --git a/drivers/cxl/core/trace.h b/drivers/cxl/core/trace.h
index c72ef9321cfe..289fab1a686d 100644
--- a/drivers/cxl/core/trace.h
+++ b/drivers/cxl/core/trace.h
@@ -7,6 +7,7 @@
 #define _CXL_EVENTS_H
 
 #include <linux/tracepoint.h>
+#include <linux/pci.h>
 #include <asm-generic/unaligned.h>
 
 #include <cxl.h>
@@ -581,6 +582,89 @@ TRACE_EVENT(cxl_memory_module,
 	)
 );
 
+#define __show_poison_source(source)                          \
+	__print_symbolic(source,                              \
+		{ CXL_POISON_SOURCE_UNKNOWN,   "Unknown"  },  \
+		{ CXL_POISON_SOURCE_EXTERNAL,  "External" },  \
+		{ CXL_POISON_SOURCE_INTERNAL,  "Internal" },  \
+		{ CXL_POISON_SOURCE_INJECTED,  "Injected" },  \
+		{ CXL_POISON_SOURCE_VENDOR,    "Vendor"   })
+
+#define show_poison_source(source)			     \
+	(((source > CXL_POISON_SOURCE_INJECTED) &&	     \
+	 (source != CXL_POISON_SOURCE_VENDOR)) ? "Reserved"  \
+	 : __show_poison_source(source))
+
+#define show_poison_flags(flags)                             \
+	__print_flags(flags, "|",                            \
+		{ CXL_POISON_FLAG_MORE,      "More"     },   \
+		{ CXL_POISON_FLAG_OVERFLOW,  "Overflow"  },  \
+		{ CXL_POISON_FLAG_SCANNING,  "Scanning"  })
+
+#define __cxl_poison_addr(record)					\
+	(le64_to_cpu(record->address))
+#define cxl_poison_record_dpa(record)					\
+	(__cxl_poison_addr(record) & CXL_POISON_START_MASK)
+#define cxl_poison_record_source(record)				\
+	(__cxl_poison_addr(record)  & CXL_POISON_SOURCE_MASK)
+#define cxl_poison_record_length(record)				\
+	(le32_to_cpu(record->length) * CXL_POISON_LEN_MULT)
+#define cxl_poison_overflow(flags, time)				\
+	(flags & CXL_POISON_FLAG_OVERFLOW ? le64_to_cpu(time) : 0)
+
+TRACE_EVENT(cxl_poison,
+
+	TP_PROTO(struct cxl_memdev *cxlmd, struct cxl_region *region,
+		 const struct cxl_poison_record *record,
+		 u8 flags, __le64 overflow_t),
+
+	TP_ARGS(cxlmd, region, record, flags, overflow_t),
+
+	TP_STRUCT__entry(
+		__string(memdev, dev_name(&cxlmd->dev))
+		__string(host, dev_name(cxlmd->dev.parent))
+		__field(u64, serial)
+		__string(region, region)
+		__field(u64, overflow_t)
+		__field(u64, dpa)
+		__field(u32, length)
+		__array(char, uuid, 16)
+		__field(u8, source)
+		__field(u8, flags)
+	    ),
+
+	TP_fast_assign(
+		__assign_str(memdev, dev_name(&cxlmd->dev));
+		__assign_str(host, dev_name(cxlmd->dev.parent));
+		__entry->serial = cxlmd->cxlds->serial;
+		__entry->overflow_t = cxl_poison_overflow(flags, overflow_t);
+		__entry->dpa = cxl_poison_record_dpa(record);
+		__entry->length = cxl_poison_record_length(record);
+		__entry->source = cxl_poison_record_source(record);
+		__entry->flags = flags;
+		if (region) {
+			__assign_str(region, dev_name(&region->dev));
+			memcpy(__entry->uuid, &region->params.uuid, 16);
+		} else {
+			__assign_str(region, "");
+			memset(__entry->uuid, 0, 16);
+		}
+	    ),
+
+	TP_printk("memdev=%s host=%s serial=%lld region=%s region_uuid=%pU dpa=0x%llx length=0x%x source=%s flags=%s overflow_time=%llu",
+		__get_str(memdev),
+		__get_str(host),
+		__entry->serial,
+		__get_str(region),
+		__entry->uuid,
+		__entry->dpa,
+		__entry->length,
+		show_poison_source(__entry->source),
+		show_poison_flags(__entry->flags),
+		__entry->overflow_t
+	)
+);
+
 #endif /* _CXL_EVENTS_H */
 
 #define TRACE_INCLUDE_FILE trace
-- 
2.37.3

[PATCH v7 3/6] cxl/memdev: Add trigger_poison_list sysfs attribute

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

When a boolean 'true' is written to this attribute the memdev driver
retrieves the poison list from the device. The list consists of
addresses that are poisoned, or would result in poison if accessed,
and the source of the poison. This attribute is only visible for
devices supporting the capability. The retrieved errors are logged
as kernel trace events with the label 'cxl_poison'.

Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 Documentation/ABI/testing/sysfs-bus-cxl | 14 +++++++
 drivers/cxl/core/memdev.c               | 56 +++++++++++++++++++++++++
 drivers/cxl/cxlmem.h                    |  2 +-
 3 files changed, 71 insertions(+), 1 deletion(-)

diff --git a/Documentation/ABI/testing/sysfs-bus-cxl b/Documentation/ABI/testing/sysfs-bus-cxl
index 3acf2f17a73f..d9421c965a3b 100644
--- a/Documentation/ABI/testing/sysfs-bus-cxl
+++ b/Documentation/ABI/testing/sysfs-bus-cxl
@@ -415,3 +415,17 @@ Description:
 		1), and checks that the hardware accepts the commit request.
 		Reading this value indicates whether the region is committed or
 		not.
+
+
+What:		/sys/bus/cxl/devices/memX/trigger_poison_list
+Date:		November, 2022
+KernelVersion:	v6.2
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) When a boolean 'true' is written to this attribute the
+		memdev driver retrieves the poison list from the device. The
+		list consists of addresses that are poisoned, or would result
+		in poison if accessed, and the source of the poison. This
+		attribute is only visible for devices supporting the
+		capability. The retrieved errors are logged as kernel
+		trace events with the label 'cxl_poison'.
diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
index 0af8856936dc..ea996057815e 100644
--- a/drivers/cxl/core/memdev.c
+++ b/drivers/cxl/core/memdev.c
@@ -106,12 +106,60 @@ static ssize_t numa_node_show(struct device *dev, struct device_attribute *attr,
 }
 static DEVICE_ATTR_RO(numa_node);
 
+static int cxl_get_poison_by_memdev(struct cxl_memdev *cxlmd)
+{
+	struct cxl_dev_state *cxlds = cxlmd->cxlds;
+	u64 offset, length;
+	int rc = 0;
+
+	/* CXL 3.0 Spec 8.2.9.8.4.1 Separate pmem and ram poison requests */
+	if (resource_size(&cxlds->pmem_res)) {
+		offset = cxlds->pmem_res.start;
+		length = resource_size(&cxlds->pmem_res);
+		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
+		if (rc)
+			return rc;
+	}
+	if (resource_size(&cxlds->ram_res)) {
+		offset = cxlds->ram_res.start;
+		length = resource_size(&cxlds->ram_res);
+		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
+		/*
+		 * Invalid Physical Address is not an error for
+		 * volatile addresses. Device support is optional.
+		 */
+		if (rc == -EFAULT)
+			rc = 0;
+	}
+	return rc;
+}
+
+static ssize_t trigger_poison_list_store(struct device *dev,
+					 struct device_attribute *attr,
+					 const char *buf, size_t len)
+{
+	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
+	bool trigger;
+	int rc;
+
+	if (kstrtobool(buf, &trigger) || !trigger)
+		return -EINVAL;
+
+	down_read(&cxl_dpa_rwsem);
+	rc = cxl_get_poison_by_memdev(cxlmd);
+	up_read(&cxl_dpa_rwsem);
+
+	return rc ? rc : len;
+}
+static DEVICE_ATTR_WO(trigger_poison_list);
+
 static struct attribute *cxl_memdev_attributes[] = {
 	&dev_attr_serial.attr,
 	&dev_attr_firmware_version.attr,
 	&dev_attr_payload_max.attr,
 	&dev_attr_label_storage_size.attr,
 	&dev_attr_numa_node.attr,
+	&dev_attr_trigger_poison_list.attr,
 	NULL,
 };
 
@@ -130,6 +178,14 @@ static umode_t cxl_memdev_visible(struct kobject *kobj, struct attribute *a,
 {
 	if (!IS_ENABLED(CONFIG_NUMA) && a == &dev_attr_numa_node.attr)
 		return 0;
+
+	if (a == &dev_attr_trigger_poison_list.attr) {
+		struct device *dev = kobj_to_dev(kobj);
+
+		if (!test_bit(CXL_MEM_COMMAND_ID_GET_POISON,
+			      to_cxl_memdev(dev)->cxlds->enabled_cmds))
+			return 0;
+	}
 	return a->mode;
 }
 
diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
index c0feb43b585d..4de15292f743 100644
--- a/drivers/cxl/cxlmem.h
+++ b/drivers/cxl/cxlmem.h
@@ -145,7 +145,7 @@ struct cxl_mbox_cmd {
 	C(FWROLLBACK, -ENXIO, "rolled back to the previous active FW"),         \
 	C(FWRESET, -ENXIO, "FW failed to activate, needs cold reset"),		\
 	C(HANDLE, -ENXIO, "one or more Event Record Handles were invalid"),     \
-	C(PADDR, -ENXIO, "physical address specified is invalid"),		\
+	C(PADDR, -EFAULT, "physical address specified is invalid"),		\
 	C(POISONLMT, -ENXIO, "poison injection limit has been reached"),        \
 	C(MEDIAFAILURE, -ENXIO, "permanent issue with the media"),		\
 	C(ABORT, -ENXIO, "background cmd was aborted by device"),               \
-- 
2.37.3

[PATCH v7 4/6] cxl/region: Provide region info to the cxl_poison trace event

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

User space may need to know which region, if any, maps the poison
address(es) logged in a cxl_poison trace event. Since the mapping
of DPAs (device physical addresses) to a region can change, the
kernel must provide this information at the time the poison list
is read. The event informs user space that at event <timestamp>
this <region> mapped to this <DPA>, which is poisoned.

The cxl_poison trace event is already wired up to log the region
name and uuid if it receives param 'struct cxl_region'.

In order to provide that cxl_region, add another method for gathering
poison - by committed endpoint decoder mappings. This method is only
available with CONFIG_CXL_REGION and is only used if a region actually
maps the memdev where poison is being read. After the poison list is
read for all the mapped resources, poison is read for the unmapped
resources, and those events are logged without the region info.

Mixed mode decoders are not currently supported in Linux. Add a debug
message to the poison request path. That will serve as an alert that
poison list retrieval needs to add support for mixed mode.

The default method remains: read the poison by memdev resource.

Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 drivers/cxl/core/core.h   |  5 +++
 drivers/cxl/core/memdev.c | 14 +++++-
 drivers/cxl/core/region.c | 89 +++++++++++++++++++++++++++++++++++++++
 3 files changed, 107 insertions(+), 1 deletion(-)

diff --git a/drivers/cxl/core/core.h b/drivers/cxl/core/core.h
index cde475e13216..4f507cb85926 100644
--- a/drivers/cxl/core/core.h
+++ b/drivers/cxl/core/core.h
@@ -25,7 +25,12 @@ void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled);
 #define CXL_DAX_REGION_TYPE(x) (&cxl_dax_region_type)
 int cxl_region_init(void);
 void cxl_region_exit(void);
+int cxl_get_poison_by_endpoint(struct device *dev, void *data);
 #else
+static inline int cxl_get_poison_by_endpoint(struct device *dev, void *data)
+{
+	return 0;
+}
 static inline void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
 {
 }
diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
index ea996057815e..c11b7bc253b4 100644
--- a/drivers/cxl/core/memdev.c
+++ b/drivers/cxl/core/memdev.c
@@ -139,14 +139,26 @@ static ssize_t trigger_poison_list_store(struct device *dev,
 					 const char *buf, size_t len)
 {
 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
+	struct cxl_port *port;
 	bool trigger;
 	int rc;
 
 	if (kstrtobool(buf, &trigger) || !trigger)
 		return -EINVAL;
 
+	port = dev_get_drvdata(&cxlmd->dev);
+	if (!port || !is_cxl_endpoint(port))
+		return -EINVAL;
+
 	down_read(&cxl_dpa_rwsem);
-	rc = cxl_get_poison_by_memdev(cxlmd);
+	if (port->commit_end == -1)
+		/* No regions mapped to this memdev */
+		rc = cxl_get_poison_by_memdev(cxlmd);
+	else
+		/* Regions mapped, collect poison by endpoint */
+		rc = device_for_each_child(&port->dev, port,
+					   cxl_get_poison_by_endpoint);
+
 	up_read(&cxl_dpa_rwsem);
 
 	return rc ? rc : len;
diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
index f29028148806..a055f8e36ef1 100644
--- a/drivers/cxl/core/region.c
+++ b/drivers/cxl/core/region.c
@@ -2213,6 +2213,95 @@ struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
 }
 EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
 
+int cxl_get_poison_by_endpoint(struct device *dev, void *data)
+{
+	struct cxl_endpoint_decoder *cxled;
+	struct cxl_port *port = data;
+	struct cxl_dev_state *cxlds;
+	struct cxl_memdev *cxlmd;
+	u64 offset, length;
+	int rc = 0;
+
+	down_read(&cxl_dpa_rwsem);
+
+	if (!is_endpoint_decoder(dev))
+		goto out;
+
+	cxled = to_cxl_endpoint_decoder(dev);
+	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
+		goto out;
+
+	/*
+	 * Regions are only created with single mode decoders: pmem or ram.
+	 * Linux does not currently support mixed mode decoders. This means
+	 * that reading poison per endpoint decoder adheres to the spec
+	 * requirement that poison reads of pmem and ram must be separated.
+	 * CXL 3.0 Spec 8.2.9.8.4.1
+	 *
+	 * Watch for future support of mixed with a dev_dbg() msg.
+	 */
+	if (cxled->mode == CXL_DECODER_MIXED) {
+		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
+		goto out;
+	}
+
+	cxlmd = cxled_to_memdev(cxled);
+	if (cxled->skip) {
+		rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);
+		if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
+			rc = 0;
+		if (rc)
+			goto out;
+	}
+	length = cxled->dpa_res->end - cxled->dpa_res->start + 1;
+	rc = cxl_mem_get_poison(cxlmd, cxled->dpa_res->start, length,
+				cxled->cxld.region);
+	if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
+		rc = 0;
+	if (rc)
+		goto out;
+
+	/* Iterate until commit_end is reached */
+	if (cxled->cxld.id < port->commit_end)
+		goto out;
+
+	/*
+	 * Reach here with the last committed decoder only.
+	 * Knowing that PMEM must always follow RAM, get poison
+	 * for unmapped ranges based on the last decoder's mode:
+	 *	ram: scan remains of ram range, then scan for pmem
+	 *	pmem: scan remains of pmem range
+	 */
+	cxlds = cxlmd->cxlds;
+
+	if (cxled->mode == CXL_DECODER_RAM) {
+		offset = cxled->dpa_res->end + 1;
+		length = resource_size(&cxlds->ram_res) - offset;
+		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
+		if (rc == -EFAULT)
+			rc = 0;
+		if (rc)
+			goto out;
+	}
+	if (cxled->mode == CXL_DECODER_PMEM) {
+		offset = cxled->dpa_res->end + 1;
+		length = resource_size(&cxlds->pmem_res) - offset;
+	} else if (resource_size(&cxlds->pmem_res)) {
+		offset = cxlds->pmem_res.start;
+		length = resource_size(&cxlds->pmem_res);
+	} else {
+		rc = 1;
+		goto out;
+	}
+	/* Final get poison call. Return rc or 1 to stop iteration. */
+	rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
+	if (!rc)
+		rc = 1;
+out:
+	up_read(&cxl_dpa_rwsem);
+	return rc;
+}
+
 static struct lock_class_key cxl_pmem_region_key;
 
 static struct cxl_pmem_region *cxl_pmem_region_alloc(struct cxl_region *cxlr)
-- 
2.37.3

[PATCH v7 5/6] cxl/trace: Add an HPA to cxl_poison trace events

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

When a cxl_poison trace event is reported for a region, the poisoned
Device Physical Address (DPA) can be translated to a Host Physical
Address (HPA) for consumption by user space.

Translate and add the resulting HPA to the cxl_poison trace event.
Follow the device decode logic as defined in the CXL Spec 3.0 Section
8.2.4.19.13.

If no region currently maps the poison, assign ULLONG_MAX to the
cxl_poison event hpa field.

Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 drivers/cxl/core/trace.c | 94 ++++++++++++++++++++++++++++++++++++++++
 drivers/cxl/core/trace.h |  9 +++-
 2 files changed, 102 insertions(+), 1 deletion(-)

diff --git a/drivers/cxl/core/trace.c b/drivers/cxl/core/trace.c
index 29ae7ce81dc5..d0403dc3c8ab 100644
--- a/drivers/cxl/core/trace.c
+++ b/drivers/cxl/core/trace.c
@@ -1,5 +1,99 @@
 // SPDX-License-Identifier: GPL-2.0-only
 /* Copyright(c) 2022 Intel Corporation. All rights reserved. */
 
+#include <cxl.h>
+#include "core.h"
+
 #define CREATE_TRACE_POINTS
 #include "trace.h"
+
+static bool cxl_is_hpa_in_range(u64 hpa, struct cxl_region *cxlr, int pos)
+{
+	struct cxl_region_params *p = &cxlr->params;
+	int gran = p->interleave_granularity;
+	int ways = p->interleave_ways;
+	u64 offset;
+
+	/* Is the hpa within this region at all */
+	if (hpa < p->res->start || hpa > p->res->end) {
+		dev_dbg(&cxlr->dev,
+			"Addr trans fail: hpa 0x%llx not in region\n", hpa);
+		return false;
+	}
+
+	/* Is the hpa in an expected chunk for its pos(-ition) */
+	offset = hpa - p->res->start;
+	offset = do_div(offset, gran * ways);
+	if ((offset >= pos * gran) && (offset < (pos + 1) * gran))
+		return true;
+
+	dev_dbg(&cxlr->dev,
+		"Addr trans fail: hpa 0x%llx not in expected chunk\n", hpa);
+
+	return false;
+}
+
+static u64 cxl_dpa_to_hpa(u64 dpa,  struct cxl_region *cxlr,
+			  struct cxl_endpoint_decoder *cxled)
+{
+	u64 dpa_offset, hpa_offset, bits_upper, mask_upper, hpa;
+	struct cxl_region_params *p = &cxlr->params;
+	int pos = cxled->pos;
+	u16 eig = 0;
+	u8 eiw = 0;
+
+	ways_to_eiw(p->interleave_ways, &eiw);
+	granularity_to_eig(p->interleave_granularity, &eig);
+
+	/*
+	 * The device position in the region interleave set was removed
+	 * from the offset at HPA->DPA translation. To reconstruct the
+	 * HPA, place the 'pos' in the offset.
+	 *
+	 * The placement of 'pos' in the HPA is determined by interleave
+	 * ways and granularity and is defined in the CXL Spec 3.0 Section
+	 * 8.2.4.19.13 Implementation Note: Device Decode Logic
+	 */
+
+	/* Remove the dpa base */
+	dpa_offset = dpa - cxl_dpa_resource_start(cxled);
+
+	mask_upper = GENMASK_ULL(51, eig + 8);
+
+	if (eiw < 8) {
+		hpa_offset = (dpa_offset & mask_upper) << eiw;
+		hpa_offset |= pos << (eig + 8);
+	} else {
+		bits_upper = (dpa_offset & mask_upper) >> (eig + 8);
+		bits_upper = bits_upper * 3;
+		hpa_offset = ((bits_upper << (eiw - 8)) + pos) << (eig + 8);
+	}
+
+	/* The lower bits remain unchanged */
+	hpa_offset |= dpa_offset & GENMASK_ULL(eig + 7, 0);
+
+	/* Apply the hpa_offset to the region base address */
+	hpa = hpa_offset + p->res->start;
+
+	if (!cxl_is_hpa_in_range(hpa, cxlr, cxled->pos))
+		return ULLONG_MAX;
+
+	return hpa;
+}
+
+u64 cxl_trace_hpa(struct cxl_region *cxlr, struct cxl_memdev *cxlmd,
+		  u64 dpa)
+{
+	struct cxl_region_params *p = &cxlr->params;
+	struct cxl_endpoint_decoder *cxled = NULL;
+
+	for (int i = 0; i <  p->nr_targets; i++) {
+		cxled = p->targets[i];
+		if (cxlmd == cxled_to_memdev(cxled))
+			break;
+	}
+	if (!cxled || cxlmd != cxled_to_memdev(cxled))
+		return ULLONG_MAX;
+
+	return cxl_dpa_to_hpa(dpa, cxlr, cxled);
+}
diff --git a/drivers/cxl/core/trace.h b/drivers/cxl/core/trace.h
index 289fab1a686d..3604f31df7a6 100644
--- a/drivers/cxl/core/trace.h
+++ b/drivers/cxl/core/trace.h
@@ -612,6 +612,8 @@ TRACE_EVENT(cxl_memory_module,
 #define cxl_poison_overflow(flags, time)				\
 	(flags & CXL_POISON_FLAG_OVERFLOW ? le64_to_cpu(time) : 0)
 
+u64 cxl_trace_hpa(struct cxl_region *cxlr, struct cxl_memdev *memdev, u64 dpa);
+
 TRACE_EVENT(cxl_poison,
 
 	TP_PROTO(struct cxl_memdev *cxlmd, struct cxl_region *region,
@@ -626,6 +628,7 @@ TRACE_EVENT(cxl_poison,
 		__field(u64, serial)
 		__string(region, region)
 		__field(u64, overflow_t)
+		__field(u64, hpa)
 		__field(u64, dpa)
 		__field(u32, length)
 		__array(char, uuid, 16)
@@ -645,18 +648,22 @@ TRACE_EVENT(cxl_poison,
 		if (region) {
 			__assign_str(region, dev_name(&region->dev));
 			memcpy(__entry->uuid, &region->params.uuid, 16);
+			__entry->hpa = cxl_trace_hpa(region, cxlmd,
+						     __entry->dpa);
 		} else {
 			__assign_str(region, "");
 			memset(__entry->uuid, 0, 16);
+			__entry->hpa = ULLONG_MAX;
 		}
 	    ),
 
-	TP_printk("memdev=%s host=%s serial=%lld region=%s region_uuid=%pU dpa=0x%llx length=0x%x source=%s flags=%s overflow_time=%llu",
+	TP_printk("memdev=%s host=%s serial=%lld region=%s region_uuid=%pU hpa=0x%llx dpa=0x%llx length=0x%x source=%s flags=%s overflow_time=%llu",
 		__get_str(memdev),
 		__get_str(host),
 		__entry->serial,
 		__get_str(region),
 		__entry->uuid,
+		__entry->hpa,
 		__entry->dpa,
 		__entry->length,
 		show_poison_source(__entry->source),
-- 
2.37.3

[PATCH v7 6/6] tools/testing/cxl: Mock support for Get Poison List

[alison.schofield]

From: Alison Schofield <alison.schofield@intel.com>

Make mock memdevs support the Get Poison List mailbox command.
Return a fake poison error record when the get poison list command
is issued.

This supports testing the kernel tracing and cxl list capabilities
for media errors.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 tools/testing/cxl/test/mem.c | 42 ++++++++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/tools/testing/cxl/test/mem.c b/tools/testing/cxl/test/mem.c
index 9263b04d35f7..2fa9c18d4c2c 100644
--- a/tools/testing/cxl/test/mem.c
+++ b/tools/testing/cxl/test/mem.c
@@ -7,6 +7,7 @@
 #include <linux/delay.h>
 #include <linux/sizes.h>
 #include <linux/bits.h>
+#include <asm/unaligned.h>
 #include <cxlmem.h>
 
 #include "trace.h"
@@ -40,6 +41,10 @@ static struct cxl_cel_entry mock_cel[] = {
 		.opcode = cpu_to_le16(CXL_MBOX_OP_GET_HEALTH_INFO),
 		.effect = cpu_to_le16(0),
 	},
+	{
+		.opcode = cpu_to_le16(CXL_MBOX_OP_GET_POISON),
+		.effect = cpu_to_le16(0),
+	},
 };
 
 /* See CXL 2.0 Table 181 Get Health Info Output Payload */
@@ -471,6 +476,8 @@ static int mock_id(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *cmd)
 			cpu_to_le64(DEV_SIZE / CXL_CAPACITY_MULTIPLIER),
 	};
 
+	put_unaligned_le24(CXL_POISON_LIST_MAX, id.poison_list_max_mer);
+
 	if (cmd->size_out < sizeof(id))
 		return -EINVAL;
 
@@ -888,6 +895,34 @@ static int mock_health_info(struct cxl_dev_state *cxlds,
 	return 0;
 }
 
+static int mock_get_poison(struct cxl_dev_state *cxlds,
+			   struct cxl_mbox_cmd *cmd)
+{
+	struct cxl_mbox_poison_payload_in *pi = cmd->payload_in;
+
+	/* Mock one poison record at pi.offset for 64 bytes */
+	struct {
+		struct cxl_mbox_poison_payload_out po;
+		struct cxl_poison_record record;
+	} __packed mock_plist = {
+		.po = {
+			.count = cpu_to_le16(1),
+		},
+		.record = {
+			.length = cpu_to_le32(1),
+			.address = cpu_to_le64(le64_to_cpu(pi->offset) +
+					       CXL_POISON_SOURCE_INJECTED),
+		},
+	};
+
+	if (cmd->size_out < sizeof(mock_plist))
+		return -EINVAL;
+
+	memcpy(cmd->payload_out, &mock_plist, sizeof(mock_plist));
+	cmd->size_out = sizeof(mock_plist);
+	return 0;
+}
+
 static int cxl_mock_mbox_send(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *cmd)
 {
 	struct device *dev = cxlds->dev;
@@ -942,6 +977,9 @@ static int cxl_mock_mbox_send(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *
 	case CXL_MBOX_OP_PASSPHRASE_SECURE_ERASE:
 		rc = mock_passphrase_secure_erase(cxlds, cmd);
 		break;
+	case CXL_MBOX_OP_GET_POISON:
+		rc = mock_get_poison(cxlds, cmd);
+		break;
 	default:
 		break;
 	}
@@ -1010,6 +1048,10 @@ static int cxl_mock_mem_probe(struct platform_device *pdev)
 	if (rc)
 		return rc;
 
+	rc = cxl_poison_state_init(cxlds);
+	if (rc)
+		return rc;
+
 	rc = cxl_dev_state_identify(cxlds);
 	if (rc)
 		return rc;
-- 
2.37.3

Re: [PATCH v7 3/6] cxl/memdev: Add trigger_poison_list sysfs attribute

[Jonathan Cameron]

On Fri, 17 Feb 2023 18:18:51 -0800
alison.schofield@intel.com wrote:

> From: Alison Schofield <alison.schofield@intel.com>
> 
> When a boolean 'true' is written to this attribute the memdev driver
> retrieves the poison list from the device. The list consists of
> addresses that are poisoned, or would result in poison if accessed,
> and the source of the poison. This attribute is only visible for
> devices supporting the capability. The retrieved errors are logged
> as kernel trace events with the label 'cxl_poison'.
> 
> Signed-off-by: Alison Schofield <alison.schofield@intel.com>

Hi Alison,

The dates and version in the docs need an update. Otherwise
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>

> ---
>  Documentation/ABI/testing/sysfs-bus-cxl | 14 +++++++
>  drivers/cxl/core/memdev.c               | 56 +++++++++++++++++++++++++
>  drivers/cxl/cxlmem.h                    |  2 +-
>  3 files changed, 71 insertions(+), 1 deletion(-)
> 
> diff --git a/Documentation/ABI/testing/sysfs-bus-cxl b/Documentation/ABI/testing/sysfs-bus-cxl
> index 3acf2f17a73f..d9421c965a3b 100644
> --- a/Documentation/ABI/testing/sysfs-bus-cxl
> +++ b/Documentation/ABI/testing/sysfs-bus-cxl
> @@ -415,3 +415,17 @@ Description:
>  		1), and checks that the hardware accepts the commit request.
>  		Reading this value indicates whether the region is committed or
>  		not.
> +
> +
> +What:		/sys/bus/cxl/devices/memX/trigger_poison_list
> +Date:		November, 2022
> +KernelVersion:	v6.2

Needs an update.

> +Contact:	linux-cxl@vger.kernel.org
> +Description:

Re: [PATCH v7 4/6] cxl/region: Provide region info to the cxl_poison trace event

[Jonathan Cameron]

On Fri, 17 Feb 2023 18:18:52 -0800
alison.schofield@intel.com wrote:

> From: Alison Schofield <alison.schofield@intel.com>
> 
> User space may need to know which region, if any, maps the poison
> address(es) logged in a cxl_poison trace event. Since the mapping
> of DPAs (device physical addresses) to a region can change, the
> kernel must provide this information at the time the poison list
> is read. The event informs user space that at event <timestamp>
> this <region> mapped to this <DPA>, which is poisoned.
> 
> The cxl_poison trace event is already wired up to log the region
> name and uuid if it receives param 'struct cxl_region'.
> 
> In order to provide that cxl_region, add another method for gathering
> poison - by committed endpoint decoder mappings. This method is only
> available with CONFIG_CXL_REGION and is only used if a region actually
> maps the memdev where poison is being read. After the poison list is
> read for all the mapped resources, poison is read for the unmapped
> resources, and those events are logged without the region info.
> 
> Mixed mode decoders are not currently supported in Linux. Add a debug
> message to the poison request path. That will serve as an alert that
> poison list retrieval needs to add support for mixed mode.
> 
> The default method remains: read the poison by memdev resource.
> 
> Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Hi Alison,

I decided to give this a test run on to of a single emulated direct
attached EP with both persistent and volatile memory, but only
part of each added to a region (so we get a skip).

Without regions it all works as expected. With them not so much - see inline

region uuid is also a bit pointless for volatile regions as I think
current code makes it all 0s.

Jonathan


> ---
>  drivers/cxl/core/core.h   |  5 +++
>  drivers/cxl/core/memdev.c | 14 +++++-
>  drivers/cxl/core/region.c | 89 +++++++++++++++++++++++++++++++++++++++
>  3 files changed, 107 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cxl/core/core.h b/drivers/cxl/core/core.h
> index cde475e13216..4f507cb85926 100644
> --- a/drivers/cxl/core/core.h
> +++ b/drivers/cxl/core/core.h
> @@ -25,7 +25,12 @@ void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled);
>  #define CXL_DAX_REGION_TYPE(x) (&cxl_dax_region_type)
>  int cxl_region_init(void);
>  void cxl_region_exit(void);
> +int cxl_get_poison_by_endpoint(struct device *dev, void *data);
>  #else
> +static inline int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> +{
> +	return 0;
> +}
>  static inline void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
>  {
>  }
> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> index ea996057815e..c11b7bc253b4 100644
> --- a/drivers/cxl/core/memdev.c
> +++ b/drivers/cxl/core/memdev.c
> @@ -139,14 +139,26 @@ static ssize_t trigger_poison_list_store(struct device *dev,
>  					 const char *buf, size_t len)
>  {
>  	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
> +	struct cxl_port *port;
>  	bool trigger;
>  	int rc;
>  
>  	if (kstrtobool(buf, &trigger) || !trigger)
>  		return -EINVAL;
>  
> +	port = dev_get_drvdata(&cxlmd->dev);
> +	if (!port || !is_cxl_endpoint(port))
> +		return -EINVAL;
> +
>  	down_read(&cxl_dpa_rwsem);
> -	rc = cxl_get_poison_by_memdev(cxlmd);
> +	if (port->commit_end == -1)
> +		/* No regions mapped to this memdev */
> +		rc = cxl_get_poison_by_memdev(cxlmd);
> +	else
> +		/* Regions mapped, collect poison by endpoint */
> +		rc = device_for_each_child(&port->dev, port,
> +					   cxl_get_poison_by_endpoint);

So this had me confused for a while.  If I setup a couple of regions
on a mixed device with a skip between (may happen anyway)
then this returns 1 which is not what we are aiming for.

Chasing it through we are reaching the last part of
cxl_get_poison_by_endpoint() which has the comment
"... Return rc or 1 to stop iteration." which is exactly 
what it is doing.

so need an if (rc == 1) rc = 0; here



> +
>  	up_read(&cxl_dpa_rwsem);
>  
>  	return rc ? rc : len;
> diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> index f29028148806..a055f8e36ef1 100644
> --- a/drivers/cxl/core/region.c
> +++ b/drivers/cxl/core/region.c
> @@ -2213,6 +2213,95 @@ struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
>  }
>  EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
>  
> +int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> +{
> +	struct cxl_endpoint_decoder *cxled;
> +	struct cxl_port *port = data;
> +	struct cxl_dev_state *cxlds;
> +	struct cxl_memdev *cxlmd;
> +	u64 offset, length;
> +	int rc = 0;
> +
> +	down_read(&cxl_dpa_rwsem);
> +
> +	if (!is_endpoint_decoder(dev))
> +		goto out;
> +
> +	cxled = to_cxl_endpoint_decoder(dev);
> +	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
> +		goto out;
> +
> +	/*
> +	 * Regions are only created with single mode decoders: pmem or ram.
> +	 * Linux does not currently support mixed mode decoders. This means
> +	 * that reading poison per endpoint decoder adheres to the spec
> +	 * requirement that poison reads of pmem and ram must be separated.
> +	 * CXL 3.0 Spec 8.2.9.8.4.1
> +	 *
> +	 * Watch for future support of mixed with a dev_dbg() msg.
> +	 */
> +	if (cxled->mode == CXL_DECODER_MIXED) {
> +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> +		goto out;
> +	}
> +
> +	cxlmd = cxled_to_memdev(cxled);
> +	if (cxled->skip) {
> +		rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);

The base of this should be the top fo the previous decoder. I'm not immediately
sure how you can get hold of that.

In my test with 1GB volatile and 1GB persistent the queries that are sent are:
0 to 0x10000000 (correct)
0 to 0x30000000 (wrong start, should be 0x10000000)

0x40000000 to 0x50000000 (correct)
0x50000000 to 0x40000000 (not correct but at least the start is correct).
See below for this one.

> +		if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> +			rc = 0;
> +		if (rc)
> +			goto out;
> +	}
> +	length = cxled->dpa_res->end - cxled->dpa_res->start + 1;
> +	rc = cxl_mem_get_poison(cxlmd, cxled->dpa_res->start, length,
> +				cxled->cxld.region);
> +	if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> +		rc = 0;
> +	if (rc)
> +		goto out;
> +
> +	/* Iterate until commit_end is reached */
> +	if (cxled->cxld.id < port->commit_end)
> +		goto out;
> +
> +	/*
> +	 * Reach here with the last committed decoder only.
> +	 * Knowing that PMEM must always follow RAM, get poison
> +	 * for unmapped ranges based on the last decoder's mode:
> +	 *	ram: scan remains of ram range, then scan for pmem
> +	 *	pmem: scan remains of pmem range
> +	 */
> +	cxlds = cxlmd->cxlds;
> +
> +	if (cxled->mode == CXL_DECODER_RAM) {
> +		offset = cxled->dpa_res->end + 1;
> +		length = resource_size(&cxlds->ram_res) - offset;
> +		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> +		if (rc == -EFAULT)
> +			rc = 0;
> +		if (rc)
> +			goto out;
> +	}
> +	if (cxled->mode == CXL_DECODER_PMEM) {
> +		offset = cxled->dpa_res->end + 1;

This is fine, unless you have a ram region in which case
the offset computed here includes that as well as the
persistent bit.

> +		length = resource_size(&cxlds->pmem_res) - offset;

So you need to take that into account with something like
- (offset - cxlds->pmem_res.start);

> +	} else if (resource_size(&cxlds->pmem_res)) {
> +		offset = cxlds->pmem_res.start;
> +		length = resource_size(&cxlds->pmem_res);
> +	} else {
> +		rc = 1;
> +		goto out;
> +	}
> +	/* Final get poison call. Return rc or 1 to stop iteration. */
> +	rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> +	if (!rc)
> +		rc = 1;
> +out:
> +	up_read(&cxl_dpa_rwsem);
> +	return rc;
> +}
> +
>  static struct lock_class_key cxl_pmem_region_key;
>  
>  static struct cxl_pmem_region *cxl_pmem_region_alloc(struct cxl_region *cxlr)

Re: [PATCH v7 4/6] cxl/region: Provide region info to the cxl_poison trace event

[Alison Schofield]

On Fri, Mar 03, 2023 at 04:46:58PM +0000, Jonathan Cameron wrote:
> On Fri, 17 Feb 2023 18:18:52 -0800
> alison.schofield@intel.com wrote:
> 
> > From: Alison Schofield <alison.schofield@intel.com>
> > 
> > User space may need to know which region, if any, maps the poison
> > address(es) logged in a cxl_poison trace event. Since the mapping
> > of DPAs (device physical addresses) to a region can change, the
> > kernel must provide this information at the time the poison list
> > is read. The event informs user space that at event <timestamp>
> > this <region> mapped to this <DPA>, which is poisoned.
> > 
> > The cxl_poison trace event is already wired up to log the region
> > name and uuid if it receives param 'struct cxl_region'.
> > 
> > In order to provide that cxl_region, add another method for gathering
> > poison - by committed endpoint decoder mappings. This method is only
> > available with CONFIG_CXL_REGION and is only used if a region actually
> > maps the memdev where poison is being read. After the poison list is
> > read for all the mapped resources, poison is read for the unmapped
> > resources, and those events are logged without the region info.
> > 
> > Mixed mode decoders are not currently supported in Linux. Add a debug
> > message to the poison request path. That will serve as an alert that
> > poison list retrieval needs to add support for mixed mode.
> > 
> > The default method remains: read the poison by memdev resource.
> > 
> > Signed-off-by: Alison Schofield <alison.schofield@intel.com>
> Hi Alison,
> 
> I decided to give this a test run on to of a single emulated direct
> attached EP with both persistent and volatile memory, but only
> part of each added to a region (so we get a skip).

This looks like issues with 'unsupported' mixed mode decoders.
How are you getting past the mixed mode decoder check?

Is there a logic error here:

A device paritioned into ram and pmem, may have:
zero or more ram decodes
  followed by zero or one skip
    followed by zero or more pmem decodes
      followed by maybe unused space

So EPs may look like these: 
ram-01, ram-02, skip-pmem-03, pmem-04
skip-pmem-01, pmem-02

But never these: (no multiple skips and only skip over ram)
ram-01, skip-ram-02, skip-pmem-03
skip-pmem-01, skip-pmem-02, 

I'm still looking at the snippets below.

Thanks,
Jonathan

> 
> Without regions it all works as expected. With them not so much - see inline
> 
> region uuid is also a bit pointless for volatile regions as I think
> current code makes it all 0s.
> 
> Jonathan
> 
> 
> > ---
> >  drivers/cxl/core/core.h   |  5 +++
> >  drivers/cxl/core/memdev.c | 14 +++++-
> >  drivers/cxl/core/region.c | 89 +++++++++++++++++++++++++++++++++++++++
> >  3 files changed, 107 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/cxl/core/core.h b/drivers/cxl/core/core.h
> > index cde475e13216..4f507cb85926 100644
> > --- a/drivers/cxl/core/core.h
> > +++ b/drivers/cxl/core/core.h
> > @@ -25,7 +25,12 @@ void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled);
> >  #define CXL_DAX_REGION_TYPE(x) (&cxl_dax_region_type)
> >  int cxl_region_init(void);
> >  void cxl_region_exit(void);
> > +int cxl_get_poison_by_endpoint(struct device *dev, void *data);
> >  #else
> > +static inline int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> > +{
> > +	return 0;
> > +}
> >  static inline void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
> >  {
> >  }
> > diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> > index ea996057815e..c11b7bc253b4 100644
> > --- a/drivers/cxl/core/memdev.c
> > +++ b/drivers/cxl/core/memdev.c
> > @@ -139,14 +139,26 @@ static ssize_t trigger_poison_list_store(struct device *dev,
> >  					 const char *buf, size_t len)
> >  {
> >  	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
> > +	struct cxl_port *port;
> >  	bool trigger;
> >  	int rc;
> >  
> >  	if (kstrtobool(buf, &trigger) || !trigger)
> >  		return -EINVAL;
> >  
> > +	port = dev_get_drvdata(&cxlmd->dev);
> > +	if (!port || !is_cxl_endpoint(port))
> > +		return -EINVAL;
> > +
> >  	down_read(&cxl_dpa_rwsem);
> > -	rc = cxl_get_poison_by_memdev(cxlmd);
> > +	if (port->commit_end == -1)
> > +		/* No regions mapped to this memdev */
> > +		rc = cxl_get_poison_by_memdev(cxlmd);
> > +	else
> > +		/* Regions mapped, collect poison by endpoint */
> > +		rc = device_for_each_child(&port->dev, port,
> > +					   cxl_get_poison_by_endpoint);
> 
> So this had me confused for a while.  If I setup a couple of regions
> on a mixed device with a skip between (may happen anyway)
> then this returns 1 which is not what we are aiming for.
> 
> Chasing it through we are reaching the last part of
> cxl_get_poison_by_endpoint() which has the comment
> "... Return rc or 1 to stop iteration." which is exactly 
> what it is doing.
> 
> so need an if (rc == 1) rc = 0; here
> 
> 
> 
> > +
> >  	up_read(&cxl_dpa_rwsem);
> >  
> >  	return rc ? rc : len;
> > diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> > index f29028148806..a055f8e36ef1 100644
> > --- a/drivers/cxl/core/region.c
> > +++ b/drivers/cxl/core/region.c
> > @@ -2213,6 +2213,95 @@ struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
> >  }
> >  EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
> >  
> > +int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> > +{
> > +	struct cxl_endpoint_decoder *cxled;
> > +	struct cxl_port *port = data;
> > +	struct cxl_dev_state *cxlds;
> > +	struct cxl_memdev *cxlmd;
> > +	u64 offset, length;
> > +	int rc = 0;
> > +
> > +	down_read(&cxl_dpa_rwsem);
> > +
> > +	if (!is_endpoint_decoder(dev))
> > +		goto out;
> > +
> > +	cxled = to_cxl_endpoint_decoder(dev);
> > +	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
> > +		goto out;
> > +
> > +	/*
> > +	 * Regions are only created with single mode decoders: pmem or ram.
> > +	 * Linux does not currently support mixed mode decoders. This means
> > +	 * that reading poison per endpoint decoder adheres to the spec
> > +	 * requirement that poison reads of pmem and ram must be separated.
> > +	 * CXL 3.0 Spec 8.2.9.8.4.1
> > +	 *
> > +	 * Watch for future support of mixed with a dev_dbg() msg.
> > +	 */
> > +	if (cxled->mode == CXL_DECODER_MIXED) {
> > +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> > +		goto out;
> > +	}
> > +
> > +	cxlmd = cxled_to_memdev(cxled);
> > +	if (cxled->skip) {
> > +		rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);
> 
> The base of this should be the top fo the previous decoder. I'm not immediately
> sure how you can get hold of that.
> 
> In my test with 1GB volatile and 1GB persistent the queries that are sent are:
> 0 to 0x10000000 (correct)
> 0 to 0x30000000 (wrong start, should be 0x10000000)
> 
> 0x40000000 to 0x50000000 (correct)
> 0x50000000 to 0x40000000 (not correct but at least the start is correct).
> See below for this one.
> 
> > +		if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> > +			rc = 0;
> > +		if (rc)
> > +			goto out;
> > +	}
> > +	length = cxled->dpa_res->end - cxled->dpa_res->start + 1;
> > +	rc = cxl_mem_get_poison(cxlmd, cxled->dpa_res->start, length,
> > +				cxled->cxld.region);
> > +	if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> > +		rc = 0;
> > +	if (rc)
> > +		goto out;
> > +
> > +	/* Iterate until commit_end is reached */
> > +	if (cxled->cxld.id < port->commit_end)
> > +		goto out;
> > +
> > +	/*
> > +	 * Reach here with the last committed decoder only.
> > +	 * Knowing that PMEM must always follow RAM, get poison
> > +	 * for unmapped ranges based on the last decoder's mode:
> > +	 *	ram: scan remains of ram range, then scan for pmem
> > +	 *	pmem: scan remains of pmem range
> > +	 */
> > +	cxlds = cxlmd->cxlds;
> > +
> > +	if (cxled->mode == CXL_DECODER_RAM) {
> > +		offset = cxled->dpa_res->end + 1;
> > +		length = resource_size(&cxlds->ram_res) - offset;
> > +		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> > +		if (rc == -EFAULT)
> > +			rc = 0;
> > +		if (rc)
> > +			goto out;
> > +	}
> > +	if (cxled->mode == CXL_DECODER_PMEM) {
> > +		offset = cxled->dpa_res->end + 1;
> 
> This is fine, unless you have a ram region in which case
> the offset computed here includes that as well as the
> persistent bit.
> 
> > +		length = resource_size(&cxlds->pmem_res) - offset;
> 
> So you need to take that into account with something like
> - (offset - cxlds->pmem_res.start);
> 
> > +	} else if (resource_size(&cxlds->pmem_res)) {
> > +		offset = cxlds->pmem_res.start;
> > +		length = resource_size(&cxlds->pmem_res);
> > +	} else {
> > +		rc = 1;
> > +		goto out;
> > +	}
> > +	/* Final get poison call. Return rc or 1 to stop iteration. */
> > +	rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> > +	if (!rc)
> > +		rc = 1;
> > +out:
> > +	up_read(&cxl_dpa_rwsem);
> > +	return rc;
> > +}
> > +
> >  static struct lock_class_key cxl_pmem_region_key;
> >  
> >  static struct cxl_pmem_region *cxl_pmem_region_alloc(struct cxl_region *cxlr)
> 

Re: [PATCH v7 3/6] cxl/memdev: Add trigger_poison_list sysfs attribute

[Alison Schofield]

On Fri, Mar 03, 2023 at 03:42:45PM +0000, Jonathan Cameron wrote:
> On Fri, 17 Feb 2023 18:18:51 -0800
> alison.schofield@intel.com wrote:
> 
> > From: Alison Schofield <alison.schofield@intel.com>
> > 
> > When a boolean 'true' is written to this attribute the memdev driver
> > retrieves the poison list from the device. The list consists of
> > addresses that are poisoned, or would result in poison if accessed,
> > and the source of the poison. This attribute is only visible for
> > devices supporting the capability. The retrieved errors are logged
> > as kernel trace events with the label 'cxl_poison'.
> > 
> > Signed-off-by: Alison Schofield <alison.schofield@intel.com>
> 
> Hi Alison,
> 
> The dates and version in the docs need an update. Otherwise
> Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>

Thanks Jonathan!

Fingers crossed for the March 2023 date!

> 
> > ---
> >  Documentation/ABI/testing/sysfs-bus-cxl | 14 +++++++
> >  drivers/cxl/core/memdev.c               | 56 +++++++++++++++++++++++++
> >  drivers/cxl/cxlmem.h                    |  2 +-
> >  3 files changed, 71 insertions(+), 1 deletion(-)
> > 
> > diff --git a/Documentation/ABI/testing/sysfs-bus-cxl b/Documentation/ABI/testing/sysfs-bus-cxl
> > index 3acf2f17a73f..d9421c965a3b 100644
> > --- a/Documentation/ABI/testing/sysfs-bus-cxl
> > +++ b/Documentation/ABI/testing/sysfs-bus-cxl
> > @@ -415,3 +415,17 @@ Description:
> >  		1), and checks that the hardware accepts the commit request.
> >  		Reading this value indicates whether the region is committed or
> >  		not.
> > +
> > +
> > +What:		/sys/bus/cxl/devices/memX/trigger_poison_list
> > +Date:		November, 2022
> > +KernelVersion:	v6.2
> 
> Needs an update.
> 
> > +Contact:	linux-cxl@vger.kernel.org
> > +Description:

Re: [PATCH v7 4/6] cxl/region: Provide region info to the cxl_poison trace event

[Alison Schofield]

On Fri, Mar 03, 2023 at 09:35:06AM -0800, Alison Schofield wrote:
> On Fri, Mar 03, 2023 at 04:46:58PM +0000, Jonathan Cameron wrote:
> > On Fri, 17 Feb 2023 18:18:52 -0800
> > alison.schofield@intel.com wrote:
> > 
> > > From: Alison Schofield <alison.schofield@intel.com>
> > > 
> > > User space may need to know which region, if any, maps the poison
> > > address(es) logged in a cxl_poison trace event. Since the mapping
> > > of DPAs (device physical addresses) to a region can change, the
> > > kernel must provide this information at the time the poison list
> > > is read. The event informs user space that at event <timestamp>
> > > this <region> mapped to this <DPA>, which is poisoned.
> > > 
> > > The cxl_poison trace event is already wired up to log the region
> > > name and uuid if it receives param 'struct cxl_region'.
> > > 
> > > In order to provide that cxl_region, add another method for gathering
> > > poison - by committed endpoint decoder mappings. This method is only
> > > available with CONFIG_CXL_REGION and is only used if a region actually
> > > maps the memdev where poison is being read. After the poison list is
> > > read for all the mapped resources, poison is read for the unmapped
> > > resources, and those events are logged without the region info.
> > > 
> > > Mixed mode decoders are not currently supported in Linux. Add a debug
> > > message to the poison request path. That will serve as an alert that
> > > poison list retrieval needs to add support for mixed mode.
> > > 
> > > The default method remains: read the poison by memdev resource.
> > > 
> > > Signed-off-by: Alison Schofield <alison.schofield@intel.com>
> > Hi Alison,
> > 
> > I decided to give this a test run on to of a single emulated direct
> > attached EP with both persistent and volatile memory, but only
> > part of each added to a region (so we get a skip).
> 
> This looks like issues with 'unsupported' mixed mode decoders.
> How are you getting past the mixed mode decoder check?

Ignore the above mixed mode decoder question.

This logic below is what I'm thinking, and coding too,
but I dont' think it's related to the issues you saw.

I've looked at each issue below, and think I've fixed.
See inline...

> 
> Is there a logic error here:
> 
> A device paritioned into ram and pmem, may have:
> zero or more ram decodes
>   followed by zero or one skip
>     followed by zero or more pmem decodes
>       followed by maybe unused space
> 
> So EPs may look like these: 
> ram-01, ram-02, skip-pmem-03, pmem-04
> skip-pmem-01, pmem-02
> 
> But never these: (no multiple skips and only skip over ram)
> ram-01, skip-ram-02, skip-pmem-03
> skip-pmem-01, skip-pmem-02, 
> 
> I'm still looking at the snippets below.
> 
> Thanks,
> Jonathan
> 
> > 
> > Without regions it all works as expected. With them not so much - see inline
> > 
> > region uuid is also a bit pointless for volatile regions as I think
> > current code makes it all 0s.

Are you referring to the cxl_poison trace_event?

The region_uuid field is being included as standard, for all cxl_poison events.
So, events without region-uuids, which includes ram regions, and plain
memdevs, will get the annoyingly region_uuid=0000000000000000000000000.

Is that bothering you ?

> > 
> > Jonathan
> > 
> > 
> > > ---
> > >  drivers/cxl/core/core.h   |  5 +++
> > >  drivers/cxl/core/memdev.c | 14 +++++-
> > >  drivers/cxl/core/region.c | 89 +++++++++++++++++++++++++++++++++++++++
> > >  3 files changed, 107 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/cxl/core/core.h b/drivers/cxl/core/core.h
> > > index cde475e13216..4f507cb85926 100644
> > > --- a/drivers/cxl/core/core.h
> > > +++ b/drivers/cxl/core/core.h
> > > @@ -25,7 +25,12 @@ void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled);
> > >  #define CXL_DAX_REGION_TYPE(x) (&cxl_dax_region_type)
> > >  int cxl_region_init(void);
> > >  void cxl_region_exit(void);
> > > +int cxl_get_poison_by_endpoint(struct device *dev, void *data);
> > >  #else
> > > +static inline int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> > > +{
> > > +	return 0;
> > > +}
> > >  static inline void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
> > >  {
> > >  }
> > > diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> > > index ea996057815e..c11b7bc253b4 100644
> > > --- a/drivers/cxl/core/memdev.c
> > > +++ b/drivers/cxl/core/memdev.c
> > > @@ -139,14 +139,26 @@ static ssize_t trigger_poison_list_store(struct device *dev,
> > >  					 const char *buf, size_t len)
> > >  {
> > >  	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
> > > +	struct cxl_port *port;
> > >  	bool trigger;
> > >  	int rc;
> > >  
> > >  	if (kstrtobool(buf, &trigger) || !trigger)
> > >  		return -EINVAL;
> > >  
> > > +	port = dev_get_drvdata(&cxlmd->dev);
> > > +	if (!port || !is_cxl_endpoint(port))
> > > +		return -EINVAL;
> > > +
> > >  	down_read(&cxl_dpa_rwsem);
> > > -	rc = cxl_get_poison_by_memdev(cxlmd);
> > > +	if (port->commit_end == -1)
> > > +		/* No regions mapped to this memdev */
> > > +		rc = cxl_get_poison_by_memdev(cxlmd);
> > > +	else
> > > +		/* Regions mapped, collect poison by endpoint */
> > > +		rc = device_for_each_child(&port->dev, port,
> > > +					   cxl_get_poison_by_endpoint);
> > 
> > So this had me confused for a while.  If I setup a couple of regions
> > on a mixed device with a skip between (may happen anyway)
> > then this returns 1 which is not what we are aiming for.
> > 
> > Chasing it through we are reaching the last part of
> > cxl_get_poison_by_endpoint() which has the comment
> > "... Return rc or 1 to stop iteration." which is exactly 
> > what it is doing.
> > 
> > so need an if (rc == 1) rc = 0; here

Thanks for the catch. I fixed up as you suggested, so that
only (-rc)'s pass onward.

> > 
> > 
> > 
> > > +
> > >  	up_read(&cxl_dpa_rwsem);
> > >  
> > >  	return rc ? rc : len;
> > > diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> > > index f29028148806..a055f8e36ef1 100644
> > > --- a/drivers/cxl/core/region.c
> > > +++ b/drivers/cxl/core/region.c
> > > @@ -2213,6 +2213,95 @@ struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
> > >  }
> > >  EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
> > >  
> > > +int cxl_get_poison_by_endpoint(struct device *dev, void *data)
> > > +{
> > > +	struct cxl_endpoint_decoder *cxled;
> > > +	struct cxl_port *port = data;
> > > +	struct cxl_dev_state *cxlds;
> > > +	struct cxl_memdev *cxlmd;
> > > +	u64 offset, length;
> > > +	int rc = 0;
> > > +
> > > +	down_read(&cxl_dpa_rwsem);
> > > +
> > > +	if (!is_endpoint_decoder(dev))
> > > +		goto out;
> > > +
> > > +	cxled = to_cxl_endpoint_decoder(dev);
> > > +	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
> > > +		goto out;
> > > +
> > > +	/*
> > > +	 * Regions are only created with single mode decoders: pmem or ram.
> > > +	 * Linux does not currently support mixed mode decoders. This means
> > > +	 * that reading poison per endpoint decoder adheres to the spec
> > > +	 * requirement that poison reads of pmem and ram must be separated.
> > > +	 * CXL 3.0 Spec 8.2.9.8.4.1
> > > +	 *
> > > +	 * Watch for future support of mixed with a dev_dbg() msg.
> > > +	 */
> > > +	if (cxled->mode == CXL_DECODER_MIXED) {
> > > +		dev_dbg(dev, "poison list read unsupported in mixed mode\n");
> > > +		goto out;
> > > +	}
> > > +
> > > +	cxlmd = cxled_to_memdev(cxled);
> > > +	if (cxled->skip) {
> > > +		rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);
> > 
> > The base of this should be the top fo the previous decoder. I'm not immediately
> > sure how you can get hold of that.

I had only consdired and coded for the case where the skip was at the
start of the cxlds resource. This diff below works because the
cxled->skip is telling us the length that was skipped over to
get to this decoder.

diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
index a055f8e36ef1..9ba7cfbdaca8 100644
--- a/drivers/cxl/core/region.c
+++ b/drivers/cxl/core/region.c
@@ -2247,15 +2247,19 @@ int cxl_get_poison_by_endpoint(struct device *dev, void *data)

        cxlmd = cxled_to_memdev(cxled);
        if (cxled->skip) {
-               rc = cxl_mem_get_poison(cxlmd, 0, cxled->skip, NULL);
+               offset = cxled->dpa_res->start - cxled->skip;
+               length = cxled->skip;
+               rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
                if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
                        rc = 0;
                if (rc)
                        goto out;
        }

Test case showing ram decode followed by a few pmem decodes:

cxl decoder8.0: cxl_get_poison_by_endpoint: endpoint: decoder8.0
cxl_mem mem0: region:region2 offset: 0x0 len: 0x10000000

cxl decoder8.1: cxl_get_poison_by_endpoint: endpoint: decoder8.1
cxl decoder8.1: cxl_get_poison_by_endpoint: SKIP FOUND
cxl_mem mem0: region:No_Region offset: 0x10000000 len: 0x30000000
(above is reading poison for the skip)
cxl_mem mem0: region:region4 offset: 0x40000000 len: 0x10000000

cxl decoder8.2: cxl_get_poison_by_endpoint: endpoint: decoder8.2
cxl_mem mem0: region:region9 offset: 0x50000000 len: 0x10000000

cxl decoder8.3: cxl_get_poison_by_endpoint: endpoint: decoder8.3
cxl_mem mem0: region:region10 offset: 0x60000000 len: 0x10000000

> > 
> > In my test with 1GB volatile and 1GB persistent the queries that are sent are:
> > 0 to 0x10000000 (correct)
> > 0 to 0x30000000 (wrong start, should be 0x10000000)
> > 
> > 0x40000000 to 0x50000000 (correct)
> > 0x50000000 to 0x40000000 (not correct but at least the start is correct).
> > See below for this one.
> > 
> > > +		if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> > > +			rc = 0;
> > > +		if (rc)
> > > +			goto out;
> > > +	}
> > > +	length = cxled->dpa_res->end - cxled->dpa_res->start + 1;
> > > +	rc = cxl_mem_get_poison(cxlmd, cxled->dpa_res->start, length,
> > > +				cxled->cxld.region);
> > > +	if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
> > > +		rc = 0;
> > > +	if (rc)
> > > +		goto out;
> > > +
> > > +	/* Iterate until commit_end is reached */
> > > +	if (cxled->cxld.id < port->commit_end)
> > > +		goto out;
> > > +
> > > +	/*
> > > +	 * Reach here with the last committed decoder only.
> > > +	 * Knowing that PMEM must always follow RAM, get poison
> > > +	 * for unmapped ranges based on the last decoder's mode:
> > > +	 *	ram: scan remains of ram range, then scan for pmem
> > > +	 *	pmem: scan remains of pmem range
> > > +	 */
> > > +	cxlds = cxlmd->cxlds;
> > > +
> > > +	if (cxled->mode == CXL_DECODER_RAM) {
> > > +		offset = cxled->dpa_res->end + 1;
> > > +		length = resource_size(&cxlds->ram_res) - offset;
> > > +		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> > > +		if (rc == -EFAULT)
> > > +			rc = 0;
> > > +		if (rc)
> > > +			goto out;
> > > +	}
> > > +	if (cxled->mode == CXL_DECODER_PMEM) {
> > > +		offset = cxled->dpa_res->end + 1;
> > 
> > This is fine, unless you have a ram region in which case
> > the offset computed here includes that as well as the
> > persistent bit.
> > 
> > > +		length = resource_size(&cxlds->pmem_res) - offset;
> > 
> > So you need to take that into account with something like
> > - (offset - cxlds->pmem_res.start);
> > 

I think this is corrected by using the cxlds->dpa_res in the
length calculation. And then, there could be no remains,
hence the length check.

        if (cxled->mode == CXL_DECODER_PMEM) {
                offset = cxled->dpa_res->end + 1;
-               length = resource_size(&cxlds->pmem_res) - offset;
+               length = resource_size(&cxlds->dpa_res) - offset;
+               if (!length) {
+                       rc = 1;
+                       goto out;
+               }


I'm going to give these changes a more thorough testing and
add more test cases before posting the next version.

Thanks for the test and review!

Alison

> > > +	} else if (resource_size(&cxlds->pmem_res)) {
> > > +		offset = cxlds->pmem_res.start;
> > > +		length = resource_size(&cxlds->pmem_res);
> > > +	} else {
> > > +		rc = 1;
> > > +		goto out;
> > > +	}
> > > +	/* Final get poison call. Return rc or 1 to stop iteration. */
> > > +	rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
> > > +	if (!rc)
> > > +		rc = 1;
> > > +out:
> > > +	up_read(&cxl_dpa_rwsem);
> > > +	return rc;
> > > +}
> > > +
> > >  static struct lock_class_key cxl_pmem_region_key;
> > >  
> > >  static struct cxl_pmem_region *cxl_pmem_region_alloc(struct cxl_region *cxlr)
> >