Skip to contentSkip to conversation

imporoved mirror DSO logging

PATCH New Malahal Naineni 2 participants 6 messages Nov 30, 2009 – Dec 8, 2009 Device mapper

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 imporoved mirror DSO logging

This patchset consists of two patches. Patch 1 just logs better error messages. Patch 2 [RFC] attempts to re-integrate a failed secondary mirror device [the second patch required a minor dm-raid1 module patch for testing].

Thanks, Malahal.

#1 lore raw

PATCH 1/1 Improve mirror DSO's failure logging

The mirror target has the following device states. The mirror DSO (daemons/dmeventd/plugins/mirror/dmeventd_mirror.c) doesn't know any of these states. This patchs adds these states to the DSO for better error reporting.

A => Alive - No failures
D => Dead - A write failure occurred leaving mirror out-of-sync S => Sync - A sychronization failure occurred, mirror out-of-sync R => Read - A read failure occurred, mirror data unaffected

daemons/dmeventd/plugins/mirror/dmeventd_mirror.c +63 -33
diff -r fff61ad560ad -r e5581203d547 daemons/dmeventd/plugins/mirror/dmeventd_mirror.c
@@ -28,9 +28,17 @@
 #include <syslog.h> /* FIXME Replace syslog with multilog */
 /* FIXME Missing openlog? */
 
-#define ME_IGNORE    0
-#define ME_INSYNC    1
-#define ME_FAILURE   2
+/* 
+ * An event may contain more than one error type. The following are bit
+ * flags that indicate each error type.
+ */
+#define ME_IGNORE    			0x01U
+#define ME_INSYNC    			0x02U
+#define ME_READ_FAILURE   		0x04U
+#define ME_SYNC_FAILURE   		0x08U
+#define ME_LOG_FAILURE 			0x10U
+#define ME_SECONDARY_WRITE_FAILURE	0x20U
+#define ME_PRIMARY_WRITE_FAILURE	0x40U
 
 /*
  * register_device() is called first and performs initialisation.
@@ -51,15 +59,16 @@ static void *_lvm_handle = NULL;
  */
 static pthread_mutex_t _event_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-static int _get_mirror_event(char *params)
+static int _get_mirror_event(const char *device, char *params)
 {
-	int i, r = ME_INSYNC;
+	int i, r;
 	char **args = NULL;
 	char *dev_status_str;
 	char *log_status_str;
 	char *sync_str;
 	char *p = NULL;
 	int log_argc, num_devs;
+	int retval = 0;
 
 	/*
 	 * dm core parms:	     0 409600 mirror
@@ -89,27 +98,48 @@ static int _get_mirror_event(char *param
 	sync_str = args[num_devs];
 
 	/* Check for bad mirror devices */
-	for (i = 0; i < num_devs; i++)
-		if (dev_status_str[i] == 'D') {
-			syslog(LOG_ERR, "Mirror device, %s, has failed.\n", args[i]);
-			r = ME_FAILURE;
+	for (i = 0; i < num_devs; i++) {
+		r = 0;
+		switch (dev_status_str[i]) {
+		case 'D': /* write failure */
+		case 'F': /* flush failure, handled as write failure */
+			syslog(LOG_ERR, "Mirror device: %s, leg: %s had a "
+					"write failure.\n", device, args[i]);
+			if (i == 0)
+				r = ME_PRIMARY_WRITE_FAILURE;
+			else 
+				r = ME_SECONDARY_WRITE_FAILURE;
+			break;
+		case 'S':
+			syslog(LOG_ERR, "Mirror device: %s, leg: %s had "
+					"sync failure.\n", device, args[i]);
+			r = ME_SYNC_FAILURE;
+			break;
+		case 'R':
+			syslog(LOG_ERR, "Mirror device: %s, leg: %s had a "
+					"read failure.\n", device, args[i]);
+			r = ME_READ_FAILURE;
+			break;
 		}
+		retval = retval | r;
+	}
 
 	/* Check for bad disk log device */
 	if (log_argc > 1 && log_status_str[0] == 'D') {
-		syslog(LOG_ERR, "Log device, %s, has failed.\n",
-		       args[2 + num_devs + log_argc]);
-		r = ME_FAILURE;
+		syslog(LOG_ERR, "Mirror device: %s, log device: %s failed.\n",
+				device, args[2 + num_devs + log_argc]);
+		retval = retval | ME_LOG_FAILURE;
 	}
 
-	if (r == ME_FAILURE)
+	if (retval)	/* A failure occurred */
 		goto out;
 
+	retval = ME_INSYNC;	/* assume INSYNC event */
 	p = strstr(sync_str, "/");
 	if (p) {
 		p[0] = '\0';
 		if (strcmp(sync_str, p+1))
-			r = ME_IGNORE;
+			retval = ME_IGNORE;
 		p[0] = '/';
 	} else
 		goto out_parse;
@@ -117,7 +147,7 @@ static int _get_mirror_event(char *param
 out:
 	if (args)
 		dm_free(args);
-	return r;
+	return retval;
 	
 out_parse:
 	if (args)
@@ -183,6 +213,7 @@ void process_event(struct dm_task *dmt,
 	char *target_type = NULL;
 	char *params;
 	const char *device = dm_task_get_name(dmt);
+	int error;
 
 	if (pthread_mutex_trylock(&_event_mutex)) {
 		syslog(LOG_NOTICE, "Another thread is handling an event.  Waiting...");
@@ -202,17 +233,11 @@ void process_event(struct dm_task *dmt,
 			continue;
 		}
 
-		switch(_get_mirror_event(params)) {
-		case ME_INSYNC:
-			/* FIXME: all we really know is that this
-			   _part_ of the device is in sync
-			   Also, this is not an error
-			*/
-			syslog(LOG_NOTICE, "%s is now in-sync\n", device);
-			break;
-		case ME_FAILURE:
-			syslog(LOG_ERR, "Device failure in %s\n", device);
-			if (_remove_failed_devices(device))
+		error = _get_mirror_event(device, params);
+		if (error & ME_LOG_FAILURE ||
+		    error & ME_PRIMARY_WRITE_FAILURE ||
+		    error & ME_SECONDARY_WRITE_FAILURE) {
+			if (_remove_failed_devices(device)) {
 				/* FIXME Why are all the error return codes unused? Get rid of them? */
 				syslog(LOG_ERR, "Failed to remove faulty devices in %s\n",
 				       device);
@@ -221,13 +246,18 @@ void process_event(struct dm_task *dmt,
 				syslog(LOG_NOTICE, "%s is now a linear device.\n",
 					device);
 			*/
-			break;
-		case ME_IGNORE:
-			break;
-		default:
-			/* FIXME Provide value then! */
-			syslog(LOG_INFO, "Unknown event received.\n");
-		}
+			}
+		} else if (error & ME_INSYNC) {
+			/* FIXME: all we really know is that this
+			   _part_ of the device is in sync
+			   Also, this is not an error
+			*/
+			syslog(LOG_NOTICE, "%s is now in-sync\n", device);
+		} else if (error & ME_READ_FAILURE ||
+			   error & ME_SYNC_FAILURE) {
+			/* Ignore these for now */
+		} else
+			syslog(LOG_INFO, "Unknown event:%u received.\n", error);
 	} while (next);
 
 	pthread_mutex_unlock(&_event_mutex);
#2 lore raw

RFC PATCH 1/1 handle a temporary secondary mirror device failure

In reply to Malahal Naineni

If "retry" policy is selected, then the mirror DSO uses "dmsetup" command to restart a resync. Right now, it is done immediately but we should do it after a configured time. Also, this is done endlessly but should be done only few configured times.

Calling 'lvchange --refresh' to re-integrate the secondary mirror upon a leg failure works if there are no device failures while the "lvchange" is running. Otherwise, the failed device is replaced with "error" target which is not what we want. For now, this patch calls "dmsetup", but should be replaced with a suitable "LVM" command.

Can the DSO call __find_config_tree_str() to get the mirror fault policy set in /etc/lvm/lvm.conf file? For now, this is hardcoded.

daemons/dmeventd/plugins/mirror/dmeventd_mirror.c +121 -9
diff -r e5581203d547 -r a0324c61e9fc daemons/dmeventd/plugins/mirror/dmeventd_mirror.c
@@ -25,6 +25,9 @@
 #include <pthread.h>
 #include <unistd.h>
 
+#include "config.h"
+#include "defaults.h"
+
 #include <syslog.h> /* FIXME Replace syslog with multilog */
 /* FIXME Missing openlog? */
 
@@ -54,6 +57,53 @@ static int _register_count = 0;
 static struct dm_pool *_mem_pool = NULL;
 static void *_lvm_handle = NULL;
 
+enum fault_policy {
+	FAULT_POLICY_INVALID,
+	FAULT_POLICY_REMOVE,
+	FAULT_POLICY_ALLOCATE,
+	FAULT_POLICY_ALLOCATE_ANYWHERE,
+	FAULT_POLICY_RETRY,
+};
+
+struct mirror_device_info {
+	enum fault_policy fault_policy;
+};
+
+static enum fault_policy fault_policy_str2enum(const char *str)
+{
+	enum fault_policy ret = FAULT_POLICY_INVALID;
+
+	if (!strcmp(str, "remove"))
+		ret = FAULT_POLICY_REMOVE;
+	else if (!strcmp(str, "allocate"))
+		ret = FAULT_POLICY_ALLOCATE;
+	else if (!strcmp(str, "allocate_anywhere"))
+		ret = FAULT_POLICY_ALLOCATE_ANYWHERE;
+	else if (!strcmp(str, "retry"))
+		ret = FAULT_POLICY_RETRY;
+
+	return ret;
+}
+
+static enum fault_policy get_mirror_fault_policy()
+{
+	const char *policy;
+	enum fault_policy ret;
+
+	/* FIXME: Malahal
+	   policy = __find_config_tree_str(NULL, "activation/mirror_device_fault_policy",
+			         DEFAULT_MIRROR_DEV_FAULT_POLICY); */
+	policy = "retry";
+	ret = fault_policy_str2enum(policy);
+	if (ret == FAULT_POLICY_INVALID) {
+		syslog(LOG_ERR, "Bad activation/mirror_device_fault_policy: "
+				"%s\n", policy);
+		ret = fault_policy_str2enum(DEFAULT_MIRROR_DEV_FAULT_POLICY);
+	}
+
+	return ret;
+}
+
 /*
  * Currently only one event can be processed at a time.
  */
@@ -168,10 +218,49 @@ static void _temporary_log_fn(int level,
 		syslog(LOG_DEBUG, "%s", format);
 }
 
+#define CMD_SIZE 256	/* FIXME Use system restriction */
+
+static int retry_failed_devices(const char *device)
+{
+	int r;
+	char cmd_str[CMD_SIZE];
+	char *vg = NULL, *lv = NULL, *layer = NULL;
+
+	if (strlen(device) > 200)  /* FIXME Use real restriction */
+		/* FIXME These return code distinctions are not used so
+		 * remove them! */
+		return -ENAMETOOLONG;
+
+	if (!dm_split_lvm_name(_mem_pool, device, &vg, &lv, &layer)) {
+		syslog(LOG_ERR, "Unable to determine VG name from %s",
+		       device);
+		/* FIXME Replace with generic error return - reason for
+		 * failure has already got logged */
+		return -ENOMEM;
+	}
+
+	/* FIXME: should be running an LVM command that is pinned.
+	 * dmsetup command may not be pinned in memory all the time.
+	 * "lvchange --refresh vg/lv" only works if there are no device
+	 * failures while it is running. Otherwise, the failed device
+	 * is replaced with "error" target which is not what we want.
+	 */
+	snprintf(cmd_str, CMD_SIZE, "dmsetup suspend %s-%s", vg, lv);
+	syslog(LOG_ERR, "Running command: %s", cmd_str);
+	r = system(cmd_str);
+	if (r == 0) {
+		snprintf(cmd_str, CMD_SIZE, "dmsetup resume %s-%s", vg, lv);
+		syslog(LOG_ERR, "Running command: %s", cmd_str);
+		r = system(cmd_str);
+	}
+
+	dm_pool_empty(_mem_pool);  /* FIXME: not safe with multiple threads */
+	return r;
+}
+
 static int _remove_failed_devices(const char *device)
 {
 	int r;
-#define CMD_SIZE 256	/* FIXME Use system restriction */
 	char cmd_str[CMD_SIZE];
 	char *vg = NULL, *lv = NULL, *layer = NULL;
 
@@ -206,7 +295,7 @@ static int _remove_failed_devices(const 
 
 void process_event(struct dm_task *dmt,
 		   enum dm_event_mask event __attribute((unused)),
-		   void **unused __attribute((unused)))
+		   void **private)
 {
 	void *next = NULL;
 	uint64_t start, length;
@@ -214,6 +303,7 @@ void process_event(struct dm_task *dmt,
 	char *params;
 	const char *device = dm_task_get_name(dmt);
 	int error;
+	struct mirror_device_info *mirror_info = *private;
 
 	if (pthread_mutex_trylock(&_event_mutex)) {
 		syslog(LOG_NOTICE, "Another thread is handling an event.  Waiting...");
@@ -236,8 +326,17 @@ void process_event(struct dm_task *dmt,
 		error = _get_mirror_event(device, params);
 		if (error & ME_LOG_FAILURE ||
 		    error & ME_PRIMARY_WRITE_FAILURE ||
-		    error & ME_SECONDARY_WRITE_FAILURE) {
-			if (_remove_failed_devices(device)) {
+		    error & ME_SECONDARY_WRITE_FAILURE ||
+		    error & ME_SYNC_FAILURE) {
+ 			if (mirror_info->fault_policy == FAULT_POLICY_RETRY &&
+					(error & ME_SECONDARY_WRITE_FAILURE ||
+					 error & ME_SYNC_FAILURE)) {
+ 				syslog(LOG_ERR, "Retrying the failed mirror "
+ 						"device.\n");
+ 				if (retry_failed_devices(device))
+ 					syslog(LOG_ERR, "Failed to refresh the "
+ 							"mirror: %s\n", device);
+ 			} else if (_remove_failed_devices(device)) {
 				/* FIXME Why are all the error return codes unused? Get rid of them? */
 				syslog(LOG_ERR, "Failed to remove faulty devices in %s\n",
 				       device);
@@ -253,9 +352,8 @@ void process_event(struct dm_task *dmt,
 			   Also, this is not an error
 			*/
 			syslog(LOG_NOTICE, "%s is now in-sync\n", device);
-		} else if (error & ME_READ_FAILURE ||
-			   error & ME_SYNC_FAILURE) {
-			/* Ignore these for now */
+		} else if (error & ME_READ_FAILURE) {
+			/* Ignore it for now */
 		} else
 			syslog(LOG_INFO, "Unknown event:%u received.\n", error);
 	} while (next);
@@ -267,9 +365,10 @@ int register_device(const char *device,
 		    const char *uuid __attribute((unused)),
 		    int major __attribute((unused)),
 		    int minor __attribute((unused)),
-		    void **unused __attribute((unused)))
+		    void **private)
 {
 	int r = 0;
+	struct mirror_device_info *mirror_info;
 
 	pthread_mutex_lock(&_register_mutex);
 
@@ -282,9 +381,19 @@ int register_device(const char *device,
 	if (!_mem_pool && !(_mem_pool = dm_pool_create("mirror_dso", 1024)))
 		goto out;
 
+	mirror_info = dm_malloc(sizeof(struct mirror_device_info));
+	if (!mirror_info) {
+		dm_pool_destroy(_mem_pool);
+		_mem_pool = NULL;
+		goto out;
+	}
+	mirror_info->fault_policy = get_mirror_fault_policy();
+	*private = mirror_info;
+
 	if (!_lvm_handle) {
 		lvm2_log_fn(_temporary_log_fn);
 		if (!(_lvm_handle = lvm2_init())) {
+			dm_free(mirror_info);
 			dm_pool_destroy(_mem_pool);
 			_mem_pool = NULL;
 			goto out;
@@ -307,8 +416,11 @@ int unregister_device(const char *device
 		      const char *uuid __attribute((unused)),
 		      int major __attribute((unused)),
 		      int minor __attribute((unused)),
-		      void **unused __attribute((unused)))
+		      void **private)
 {
+	struct mirror_device_info *mirror_info = *private;
+
+	dm_free(mirror_info);
 	pthread_mutex_lock(&_register_mutex);
 
 	syslog(LOG_INFO, "No longer monitoring mirror device %s for events\n",
doc/example.conf +3 -0
diff -r e5581203d547 -r a0324c61e9fc doc/example.conf
@@ -392,6 +392,9 @@ activation {
     #            since it would break the redundant nature of the mirror. This
     #            policy acts like "remove" if no suitable device and space can
     #            be allocated for the replacement.
+    #
+    # "retry" - Try to re-integrate the failed mirror leg assuming that the
+    #           failure is transient. Not implemented yet, so don't use it.
 
     mirror_log_fault_policy = "allocate"
     mirror_device_fault_policy = "remove"
lib/metadata/mirror.c +3 -0
diff -r e5581203d547 -r a0324c61e9fc lib/metadata/mirror.c
@@ -36,6 +36,7 @@
 #define MIRROR_REMOVE		 0
 #define MIRROR_ALLOCATE		 1
 #define MIRROR_ALLOCATE_ANYWHERE 2
+#define MIRROR_RETRY	 	 3
 
 /*
  * Returns true if the lv is temporary mirror layer for resync
@@ -773,6 +774,8 @@ static int get_mirror_fault_policy(struc
 		return MIRROR_ALLOCATE;
 	else if (!strcmp(policy, "allocate_anywhere"))
 		return MIRROR_ALLOCATE_ANYWHERE;
+	else if (!strcmp(policy, "retry"))
+		return MIRROR_RETRY;
 
 	if (log_policy)
 		log_error("Bad activation/mirror_log_fault_policy");
#3 lore raw

RFC PATCH 1/1 handle a temporary secondary mirror device failure

Quoted 4 lines · Malahal Naineni

If "retry" policy is selected, then the mirror DSO uses "dmsetup" command to restart a resync. Right now, it is done immediately but we should do it after a configured time. Also, this is done endlessly but should be done only few configured times.

It sounds good to me. This feature would work in a variety of environments by adding configurable parameter as you mentioned.

Review

Quoted 8 lines

Calling 'lvchange --refresh' to re-integrate the secondary mirror upon a leg failure works if there are no device failures while the "lvchange" is running. Otherwise, the failed device is replaced with "error" target which is not what we want. For now, this patch calls "dmsetup", but should be replaced with a suitable "LVM" command.

(quoted diff) +2 -0
+	snprintf(cmd_str, CMD_SIZE, "dmsetup suspend %s-%s", vg, lv);
+	syslog(LOG_ERR, "Running command: %s", cmd_str);

We need to re-queue blocked I/Os instead of finishing them by -EIO. I think that "--noflush" option is necessary for "dmsetup suspend."

Thanks,
Taka

#4 lore raw

RFC PATCH 1/1 handle a temporary secondary mirror device failure

Quoted 15 lines · Takahiro Yasui [tyasui at redhat.com]
Quoted 4 lines · Malahal Naineni

If "retry" policy is selected, then the mirror DSO uses "dmsetup" command to restart a resync. Right now, it is done immediately but we should do it after a configured time. Also, this is done endlessly but should be done only few configured times.

It sounds good to me. This feature would work in a variety of environments by adding configurable parameter as you mentioned.

Quoted 5 lines

Calling 'lvchange --refresh' to re-integrate the secondary mirror upon a leg failure works if there are no device failures while the "lvchange" is running. Otherwise, the failed device is replaced with "error" target which is not what we want. For now, this patch calls "dmsetup", but should be replaced with a suitable "LVM" command.

Quoted 5 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+	snprintf(cmd_str, CMD_SIZE, "dmsetup suspend %s-%s", vg, lv);
+	syslog(LOG_ERR, "Running command: %s", cmd_str);

We need to re-queue blocked I/Os instead of finishing them by -EIO. I think that "--noflush" option is necessary for "dmsetup suspend."

Yes, you are right. I tested this with existing code dm-raid1 module which doesn't hold any I/O's, so it worked fine. But I think we should use some LVM command (or create one if such a command doesn't exist) instead of "dmsetup" command.

Thank you for looking into it.

Thanks, Malahal.

#5 lore raw

RFC PATCH 1/1 handle a temporary secondary mirror device failure

Quoted 22 lines · malahal at us.ibm.com
Quoted 15 lines · Takahiro Yasui [tyasui at redhat.com]
Quoted 4 lines · Malahal Naineni

If "retry" policy is selected, then the mirror DSO uses "dmsetup" command to restart a resync. Right now, it is done immediately but we should do it after a configured time. Also, this is done endlessly but should be done only few configured times.

It sounds good to me. This feature would work in a variety of environments by adding configurable parameter as you mentioned.

Quoted 5 lines

Calling 'lvchange --refresh' to re-integrate the secondary mirror upon a leg failure works if there are no device failures while the "lvchange" is running. Otherwise, the failed device is replaced with "error" target which is not what we want. For now, this patch calls "dmsetup", but should be replaced with a suitable "LVM" command.

Quoted 5 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+	snprintf(cmd_str, CMD_SIZE, "dmsetup suspend %s-%s", vg, lv);
+	syslog(LOG_ERR, "Running command: %s", cmd_str);

We need to re-queue blocked I/Os instead of finishing them by -EIO. I think that "--noflush" option is necessary for "dmsetup suspend."

Quoted 4 lines

Yes, you are right. I tested this with existing code dm-raid1 module which doesn't hold any I/O's, so it worked fine. But I think we should use some LVM command (or create one if such a command doesn't exist) instead of "dmsetup" command.

lvm commands have an unwelcome behavior to scan all devices except for ones filtered out by lvm configuration as I described before.

https://www.redhat.com/archives/lvm-devel/2009-April/msg00014.html

If "refresh" is just an action in the device-mapper layer, I rather hope that it will be implemented without lvm commands as you did. It would avoid extra time by issuing unnecessary I/Os and achieve quick recovery.

It is best that this device scanning issue of lvm commands is solved. :)

Thanks,
Taka