zed: Control NVMe fault LEDs

The ZED code currently can only turn on the fault LED for
a faulted disk in a JBOD enclosure.  This extends support
for faulted NVMe disks as well.

Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Tony Hutter <[email protected]>
Closes #12648
Closes #12695

5 files changed, 229 insertions, 11 deletions

diff --git a/cmd/zed/zed.d/statechange-led.sh b/cmd/zed/zed.d/statechange-led.sh
index 0f9da3204..26e6064fa 100755
--- a/cmd/zed/zed.d/statechange-led.sh
+++ b/cmd/zed/zed.d/statechange-led.sh
@@ -29,7 +29,8 @@
 [ -f "${ZED_ZEDLET_DIR}/zed.rc" ] && . "${ZED_ZEDLET_DIR}/zed.rc"
 . "${ZED_ZEDLET_DIR}/zed-functions.sh"
 
-if [ ! -d /sys/class/enclosure ] ; then
+if [ ! -d /sys/class/enclosure ] && [ ! -d /sys/bus/pci/slots ] ; then
+	# No JBOD enclosure or NVMe slots
 	exit 1
 fi
 
@@ -92,6 +93,29 @@ check_and_set_led()
 	done
 }
 
+# Fault LEDs for JBODs and NVMe drives are handled a little differently.
+#
+# On JBODs the fault LED is called 'fault' and on a path like this:
+#
+#   /sys/class/enclosure/0:0:1:0/SLOT 10/fault
+#
+# On NVMe it's called 'attention' and on a path like this:
+#
+#   /sys/bus/pci/slot/0/attention
+#
+# This function returns the full path to the fault LED file for a given
+# enclosure/slot directory.
+#
+path_to_led()
+{
+	dir=$1
+	if [ -f "$dir/fault" ] ; then
+		echo "$dir/fault"
+	elif [ -f "$dir/attention" ] ; then
+		echo "$dir/attention"
+	fi
+}
+
 state_to_val()
 {
 	state="$1"
@@ -105,6 +129,38 @@ state_to_val()
 	esac
 }
 
+#
+# Given a nvme name like 'nvme0n1', pass back its slot directory
+# like "/sys/bus/pci/slots/0"
+#
+nvme_dev_to_slot()
+{
+	dev="$1"
+
+	# Get the address "0000:01:00.0"
+	address=$(cat "/sys/class/block/$dev/device/address")
+
+	# For each /sys/bus/pci/slots subdir that is an actual number
+	# (rather than weird directories like "1-3/").
+	# shellcheck disable=SC2010
+	for i in $(ls /sys/bus/pci/slots/ | grep -E "^[0-9]+$") ; do
+		this_address=$(cat "/sys/bus/pci/slots/$i/address")
+
+		# The format of address is a little different between
+		# /sys/class/block/$dev/device/address and
+		# /sys/bus/pci/slots/
+		#
+		# address=           "0000:01:00.0"
+		# this_address =     "0000:01:00"
+		#
+		if echo "$address" | grep -Eq ^"$this_address" ; then
+			echo "/sys/bus/pci/slots/$i"
+			break
+		fi
+	done
+}
+
+
 # process_pool (pool)
 #
 # Iterate through a pool and set the vdevs' enclosure slot LEDs to
@@ -134,6 +190,11 @@ process_pool()
 		# Get dev name (like 'sda')
 		dev=$(basename "$(echo "$therest" | awk '{print $(NF-1)}')")
 		vdev_enc_sysfs_path=$(realpath "/sys/class/block/$dev/device/enclosure_device"*)
+		if [ ! -d "$vdev_enc_sysfs_path" ] ; then
+			# This is not a JBOD disk, but it could be a PCI NVMe drive
+			vdev_enc_sysfs_path=$(nvme_dev_to_slot "$dev")
+		fi
+
 		current_val=$(echo "$therest" | awk '{print $NF}')
 
 		if [ "$current_val" != "0" ] ; then
@@ -145,9 +206,10 @@ process_pool()
 			continue
 		fi
 
-		if [ ! -e "$vdev_enc_sysfs_path/fault" ] ; then
+		led_path=$(path_to_led "$vdev_enc_sysfs_path")
+		if [ ! -e "$led_path" ] ; then
 			rc=3
-			zed_log_msg "vdev $vdev '$file/fault' doesn't exist"
+			zed_log_msg "vdev $vdev '$led_path' doesn't exist"
 			continue
 		fi
 
@@ -158,7 +220,7 @@ process_pool()
 			continue
 		fi
 
-		if ! check_and_set_led "$vdev_enc_sysfs_path/fault" "$val"; then
+		if ! check_and_set_led "$led_path" "$val"; then
 			rc=3
 		fi
 	done
@@ -169,7 +231,8 @@ if [ -n "$ZEVENT_VDEV_ENC_SYSFS_PATH" ] && [ -n "$ZEVENT_VDEV_STATE_STR" ] ; the
 	# Got a statechange for an individual vdev
 	val=$(state_to_val "$ZEVENT_VDEV_STATE_STR")
 	vdev=$(basename "$ZEVENT_VDEV_PATH")
-	check_and_set_led "$ZEVENT_VDEV_ENC_SYSFS_PATH/fault" "$val"
+	ledpath=$(path_to_led "$ZEVENT_VDEV_ENC_SYSFS_PATH")
+	check_and_set_led "$ledpath" "$val"
 else
 	# Process the entire pool
 	poolname=$(zed_guid_to_pool "$ZEVENT_POOL_GUID")
diff --git a/cmd/zed/zed.d/zed.rc b/cmd/zed/zed.d/zed.rc
index df560f921..1c278b2ef 100644
--- a/cmd/zed/zed.d/zed.rc
+++ b/cmd/zed/zed.d/zed.rc
@@ -89,8 +89,8 @@
 
 ##
 # Turn on/off enclosure LEDs when drives get DEGRADED/FAULTED.  This works for
-# device mapper and multipath devices as well.  Your enclosure must be
-# supported by the Linux SES driver for this to work.
+# device mapper and multipath devices as well.  This works with JBOD enclosures
+# and NVMe PCI drives (assuming they're supported by Linux in sysfs).
 #
 ZED_USE_ENCLOSURE_LEDS=1
 
diff --git a/cmd/zpool/zpool.d/ses b/cmd/zpool/zpool.d/ses
index f6b7520df..b1836d676 100755
--- a/cmd/zpool/zpool.d/ses
+++ b/cmd/zpool/zpool.d/ses
@@ -41,7 +41,13 @@ for i in $scripts ; do
 		val=$(ls "$VDEV_ENC_SYSFS_PATH/../device/scsi_generic" 2>/dev/null)
 		;;
 	fault_led)
-		val=$(cat "$VDEV_ENC_SYSFS_PATH/fault" 2>/dev/null)
+		# JBODs fault LED is called 'fault', NVMe fault LED is called
+		# 'attention'.
+		if [ -f "$VDEV_ENC_SYSFS_PATH/fault" ] ; then
+			val=$(cat "$VDEV_ENC_SYSFS_PATH/fault" 2>/dev/null)
+		elif [ -f "$VDEV_ENC_SYSFS_PATH/attention" ] ; then
+			val=$(cat "$VDEV_ENC_SYSFS_PATH/attention" 2>/dev/null)
+		fi
 		;;
 	locate_led)
 		val=$(cat "$VDEV_ENC_SYSFS_PATH/locate" 2>/dev/null)
diff --git a/lib/libzutil/os/linux/zutil_device_path_os.c b/lib/libzutil/os/linux/zutil_device_path_os.c
index 2a6f4ae2a..13f8bd031 100644
--- a/lib/libzutil/os/linux/zutil_device_path_os.c
+++ b/lib/libzutil/os/linux/zutil_device_path_os.c
@@ -155,17 +155,147 @@ zfs_strip_path(char *path)
 }
 
 /*
+ * Read the contents of a sysfs file into an allocated buffer and remove the
+ * last newline.
+ *
+ * This is useful for reading sysfs files that return a single string.  Return
+ * an allocated string pointer on success, NULL otherwise.  Returned buffer
+ * must be freed by the user.
+ */
+static char *
+zfs_read_sysfs_file(char *filepath)
+{
+	char buf[4096];	/* all sysfs files report 4k size */
+	char *str = NULL;
+
+	FILE *fp = fopen(filepath, "r");
+	if (fp == NULL) {
+		return (NULL);
+	}
+	if (fgets(buf, sizeof (buf), fp) == buf) {
+		/* success */
+
+		/* Remove the last newline (if any) */
+		size_t len = strlen(buf);
+		if (buf[len - 1] == '\n') {
+			buf[len - 1] = '\0';
+		}
+		str = strdup(buf);
+	}
+
+	fclose(fp);
+
+	return (str);
+}
+
+/*
+ * Given a dev name like "nvme0n1", return the full PCI slot sysfs path to
+ * the drive (in /sys/bus/pci/slots).
+ *
+ * For example:
+ *     dev:            "nvme0n1"
+ *     returns:        "/sys/bus/pci/slots/0"
+ *
+ * 'dev' must be an NVMe device.
+ *
+ * Returned string must be freed.  Returns NULL on error or no sysfs path.
+ */
+static char *
+zfs_get_pci_slots_sys_path(const char *dev_name)
+{
+	DIR *dp = NULL;
+	struct dirent *ep;
+	char *address1 = NULL;
+	char *address2 = NULL;
+	char *path = NULL;
+	char buf[MAXPATHLEN];
+	char *tmp;
+
+	/* If they preface 'dev' with a path (like "/dev") then strip it off */
+	tmp = strrchr(dev_name, '/');
+	if (tmp != NULL)
+		dev_name = tmp + 1;    /* +1 since we want the chr after '/' */
+
+	if (strncmp("nvme", dev_name, 4) != 0)
+		return (NULL);
+
+	(void) snprintf(buf, sizeof (buf), "/sys/block/%s/device/address",
+	    dev_name);
+
+	address1 = zfs_read_sysfs_file(buf);
+	if (!address1)
+		return (NULL);
+
+	/*
+	 * /sys/block/nvme0n1/device/address format will
+	 * be "0000:01:00.0" while /sys/bus/pci/slots/0/address will be
+	 * "0000:01:00".  Just NULL terminate at the '.' so they match.
+	 */
+	tmp = strrchr(address1, '.');
+	if (tmp != NULL)
+		*tmp = '\0';
+
+	dp = opendir("/sys/bus/pci/slots/");
+	if (dp == NULL) {
+		free(address1);
+		return (NULL);
+	}
+
+	/*
+	 * Look through all the /sys/bus/pci/slots/ subdirs
+	 */
+	while ((ep = readdir(dp))) {
+		/*
+		 * We only care about directory names that are a single number.
+		 * Sometimes there's other directories like
+		 * "/sys/bus/pci/slots/0-3/" in there - skip those.
+		 */
+		if (!zfs_isnumber(ep->d_name))
+			continue;
+
+		(void) snprintf(buf, sizeof (buf),
+		    "/sys/bus/pci/slots/%s/address", ep->d_name);
+
+		address2 = zfs_read_sysfs_file(buf);
+		if (!address2)
+			continue;
+
+		if (strcmp(address1, address2) == 0) {
+			/* Addresses match, we're all done */
+			free(address2);
+			if (asprintf(&path, "/sys/bus/pci/slots/%s",
+			    ep->d_name) == -1) {
+				free(tmp);
+				continue;
+			}
+			break;
+		}
+		free(address2);
+	}
+
+	closedir(dp);
+	free(address1);
+
+	return (path);
+}
+
+/*
  * Given a dev name like "sda", return the full enclosure sysfs path to
  * the disk.  You can also pass in the name with "/dev" prepended
- * to it (like /dev/sda).
+ * to it (like /dev/sda).  This works for both JBODs and NVMe PCI devices.
  *
  * For example, disk "sda" in enclosure slot 1:
- *     dev:            "sda"
+ *     dev_name:       "sda"
  *     returns:        "/sys/class/enclosure/1:0:3:0/Slot 1"
  *
+ * Or:
+ *
+ *      dev_name:   "nvme0n1"
+ *      returns:    "/sys/bus/pci/slots/0"
+ *
  * 'dev' must be a non-devicemapper device.
  *
- * Returned string must be freed.
+ * Returned string must be freed.  Returns NULL on error.
  */
 char *
 zfs_get_enclosure_sysfs_path(const char *dev_name)
@@ -252,6 +382,16 @@ end:
 	if (dp != NULL)
 		closedir(dp);
 
+	if (!path) {
+		/*
+		 * This particular disk isn't in a JBOD.  It could be an NVMe
+		 * drive. If so, look up the NVMe device's path in
+		 * /sys/bus/pci/slots/. Within that directory is a 'attention'
+		 * file which controls the NVMe fault LED.
+		 */
+		path = zfs_get_pci_slots_sys_path(dev_name);
+	}
+
 	return (path);
 }
 
diff --git a/lib/libzutil/zutil_nicenum.c b/lib/libzutil/zutil_nicenum.c
index 1a19db0df..4dcac1f85 100644
--- a/lib/libzutil/zutil_nicenum.c
+++ b/lib/libzutil/zutil_nicenum.c
@@ -27,6 +27,7 @@
 #include <math.h>
 #include <stdio.h>
 #include <libzutil.h>
+#include <string.h>
 
 /*
  * Return B_TRUE if "str" is a number string, B_FALSE otherwise.
@@ -42,6 +43,14 @@ zfs_isnumber(const char *str)
 		if (!(isdigit(*str) || (*str == '.')))
 			return (B_FALSE);
 
+	/*
+	 * Numbers should not end with a period ("." ".." or "5." are
+	 * not valid)
+	 */
+	if (str[strlen(str) - 1] == '.') {
+		return (B_FALSE);
+	}
+
 	return (B_TRUE);
 }