Skip to contentSkip to conversation

fix ->run failure handling

PATCH In review Christoph Hellwig 3 participants 9 messages Jun 4, 2024 – Jun 12, 2024 MD RAID

lore
Share & tools
lore
#0 lore raw

Hi Song, hi Yu,

this series fixes problems I found when testing how md handles trying to creat an array with disks that support incompatible protection information or a mix of protection information and not protection information.

Note that this only fixes the oops due to the double frees. After the attempt to at the incompatible device the /dev/md0 node is still around with zero blocks and linked to the the previous devices in sysfs which is a bit confusing, but unrelated to the resource cleanup touched here.

raid10 and raid5 also free conf on failure in their ->run handle, but set mddev->private to NULL which prevents the double free as well (but requires more code)

Diffstat:

 raid0.c |   21 +++++----------------
 raid1.c |   14 +++-----------
 2 files changed, 8 insertions(+), 27 deletions(-)
#1 lore raw

PATCH 1/2 md/raid0: don't free conf on raid0_run failure

The core md code calls the ->free method which already frees conf.

Fixes 0c031fd37f69 ("md: Move alloc/free acct bioset in to personality")
---
 drivers/md/raid0.c | 21 +++++----------------
 1 file changed, 5 insertions(+), 16 deletions(-)
drivers/md/raid0.c +5 -16
index c5d4aeb68404c9..81c01347cd24e6 100644
@@ -365,18 +365,13 @@ static sector_t raid0_size(struct mddev *mddev, sector_t sectors, int raid_disks
 	return array_sectors;
 }
 
-static void free_conf(struct mddev *mddev, struct r0conf *conf)
-{
-	kfree(conf->strip_zone);
-	kfree(conf->devlist);
-	kfree(conf);
-}
-
 static void raid0_free(struct mddev *mddev, void *priv)
 {
 	struct r0conf *conf = priv;
 
-	free_conf(mddev, conf);
+	kfree(conf->strip_zone);
+	kfree(conf->devlist);
+	kfree(conf);
 }
 
 static int raid0_set_limits(struct mddev *mddev)
@@ -415,7 +410,7 @@ static int raid0_run(struct mddev *mddev)
 	if (!mddev_is_dm(mddev)) {
 		ret = raid0_set_limits(mddev);
 		if (ret)
-			goto out_free_conf;
+			return ret;
 	}
 
 	/* calculate array device size */
@@ -427,13 +422,7 @@ static int raid0_run(struct mddev *mddev)
 
 	dump_zones(mddev);
 
-	ret = md_integrity_register(mddev);
-	if (ret)
-		goto out_free_conf;
-	return 0;
-out_free_conf:
-	free_conf(mddev, conf);
-	return ret;
+	return md_integrity_register(mddev);
 }
 
 /*
Signature · 1 line
2.43.0
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] md/raid0: don't free conf on raid0_run failure

Quoted 4 lines · Christoph Hellwig

The core md code calls the ->free method which already frees conf.

Fixes 0c031fd37f69 ("md: Move alloc/free acct bioset in to personality")

LGTM

Quoted 55 lines
---
  drivers/md/raid0.c | 21 +++++----------------
  1 file changed, 5 insertions(+), 16 deletions(-)
drivers/md/raid0.c +5 -16
index c5d4aeb68404c9..81c01347cd24e6 100644
@@ -365,18 +365,13 @@ static sector_t raid0_size(struct mddev *mddev, sector_t sectors, int raid_disks
  	return array_sectors;
  }
 
-static void free_conf(struct mddev *mddev, struct r0conf *conf)
-{
-	kfree(conf->strip_zone);
-	kfree(conf->devlist);
-	kfree(conf);
-}
-
  static void raid0_free(struct mddev *mddev, void *priv)
  {
  	struct r0conf *conf = priv;
 
-	free_conf(mddev, conf);
+	kfree(conf->strip_zone);
+	kfree(conf->devlist);
+	kfree(conf);
  }
 
  static int raid0_set_limits(struct mddev *mddev)
@@ -415,7 +410,7 @@ static int raid0_run(struct mddev *mddev)
  	if (!mddev_is_dm(mddev)) {
  		ret = raid0_set_limits(mddev);
  		if (ret)
-			goto out_free_conf;
+			return ret;
  	}
 
  	/* calculate array device size */
@@ -427,13 +422,7 @@ static int raid0_run(struct mddev *mddev)
 
  	dump_zones(mddev);
 
-	ret = md_integrity_register(mddev);
-	if (ret)
-		goto out_free_conf;
-	return 0;
-out_free_conf:
-	free_conf(mddev, conf);
-	return ret;
+	return md_integrity_register(mddev);
  }
 
  /*
#3 lore raw

Re: [PATCH 1/2] md/raid0: don't free conf on raid0_run failure

Hi,

Quoted 8 lines · Yu Kuai
Quoted 4 lines · Christoph Hellwig

The core md code calls the ->free method which already frees conf.

Fixes 0c031fd37f69 ("md: Move alloc/free acct bioset in to personality")

LGTM

Sorry that I just noticed that the sysfs api level_store() calls the pers->run() as well, and it didn't handle failure by pers->free(). It's weried that the api will return success in this case, and after this patch, it will require __md_stop() to free the conf.

Can you also fix the level_store()? By checking pers->run() and if it failed, call pers->free() and return error.

Thanks,
Kuai

Quoted 53 lines
Quoted 40 lines

---
  drivers/md/raid0.c | 21 +++++----------------

  1 file changed, 5 insertions(+), 16 deletions(-)
drivers/md/raid0.c +0 -0
index c5d4aeb68404c9..81c01347cd24e6 100644
@@ -365,18 +365,13 @@ static sector_t raid0_size(struct mddev *mddev, 

sector_t sectors, int raid_disks
      return array_sectors;
  }

drivers/md/raid0.c +0 -7
-static void free_conf(struct mddev *mddev, struct r0conf *conf)
-{
-    kfree(conf->strip_zone);
-    kfree(conf->devlist);
-    kfree(conf);
-}
-

  static void raid0_free(struct mddev *mddev, void *priv) {
      struct r0conf *conf = priv;
-    free_conf(mddev, conf);
+    kfree(conf->strip_zone);
+    kfree(conf->devlist);
+    kfree(conf);
  }
  static int raid0_set_limits(struct mddev *mddev)

drivers/md/raid0.c +1 -2
@@ -415,7 +410,7 @@ static int raid0_run(struct mddev *mddev)
       if (!mddev_is_dm(mddev)) {
           ret = raid0_set_limits(mddev);
           if (ret)
-            goto out_free_conf;
+            return ret;
       }
       /* calculate array device size */
@@ -427,13 +422,7 @@ static int raid0_run(struct mddev *mddev)
       dump_zones(mddev);
-    ret = md_integrity_register(mddev);

drivers/md/raid0.c static int raid0_run(struct mddev *mddev) · line 426

Quoted 10 lines
drivers/md/raid0.c +1 -6
-    if (ret)
-        goto out_free_conf;
-    return 0;
-out_free_conf:
-    free_conf(mddev, conf);
-    return ret;
+    return md_integrity_register(mddev);
   }
   /*
 

.

#4 lore raw

Re: [PATCH 1/2] md/raid0: don't free conf on raid0_run failure

Quoted 7 lines · Yu Kuai

Sorry that I just noticed that the sysfs api level_store() calls the pers->run() as well, and it didn't handle failure by pers->free(). It's weried that the api will return success in this case, and after this patch, it will require __md_stop() to free the conf.

Can you also fix the level_store()? By checking pers->run() and if it failed, call pers->free() and return error.

I can look into it. But if we don't have consistent callers anyway I'd be tempted to not call ->free on ->run failure, as that is a rather unusual convention.

#5 lore raw

PATCH 2/2 md/raid1: don't free conf on raid0_run failure

In reply to Christoph Hellwig

The core md code calls the ->free method which already frees conf.

Fixes 07f1a6850c5d ("md/raid1: fail run raid1 array when active disk less than one")
---
 drivers/md/raid1.c | 14 +++-----------
 1 file changed, 3 insertions(+), 11 deletions(-)
drivers/md/raid1.c +3 -11
index 7b8a71ca66dde0..1f321826ef02ba 100644
@@ -3204,7 +3204,6 @@ static int raid1_set_limits(struct mddev *mddev)
 	return queue_limits_set(mddev->gendisk->queue, &lim);
 }
 
-static void raid1_free(struct mddev *mddev, void *priv);
 static int raid1_run(struct mddev *mddev)
 {
 	struct r1conf *conf;
@@ -3238,7 +3237,7 @@ static int raid1_run(struct mddev *mddev)
 	if (!mddev_is_dm(mddev)) {
 		ret = raid1_set_limits(mddev);
 		if (ret)
-			goto abort;
+			return ret;
 	}
 
 	mddev->degraded = 0;
@@ -3252,8 +3251,7 @@ static int raid1_run(struct mddev *mddev)
 	 */
 	if (conf->raid_disks - mddev->degraded < 1) {
 		md_unregister_thread(mddev, &conf->thread);
-		ret = -EINVAL;
-		goto abort;
+		return -EINVAL;
 	}
 
 	if (conf->raid_disks - mddev->degraded == 1)
@@ -3277,14 +3275,8 @@ static int raid1_run(struct mddev *mddev)
 	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
 
 	ret = md_integrity_register(mddev);
-	if (ret) {
+	if (ret)
 		md_unregister_thread(mddev, &mddev->thread);
-		goto abort;
-	}
-	return 0;
-
-abort:
-	raid1_free(mddev, conf);
 	return ret;
 }
 
Signature · 1 line
2.43.0
maintainer
#6 lore raw
In reply to Christoph Hellwig

Hi Christoph,

Quoted 14 lines · Christoph Hellwig

Hi Song, hi Yu,

this series fixes problems I found when testing how md handles trying to creat an array with disks that support incompatible protection information or a mix of protection information and not protection information.

Note that this only fixes the oops due to the double frees. After the attempt to at the incompatible device the /dev/md0 node is still around with zero blocks and linked to the the previous devices in sysfs which is a bit confusing, but unrelated to the resource cleanup touched here.

Quoted 8 lines

raid10 and raid5 also free conf on failure in their ->run handle, but set mddev->private to NULL which prevents the double free as well (but requires more code)

Diffstat:

 raid0.c |   21 +++++----------------
 raid1.c |   14 +++-----------
 2 files changed, 8 insertions(+), 27 deletions(-)

Thanks for the patchset. I applied it to the md-6.11 branch.

Song

#7 lore raw
Quoted 10 lines · Song Liu
Quoted 8 lines

raid10 and raid5 also free conf on failure in their ->run handle, but set mddev->private to NULL which prevents the double free as well (but requires more code)

Diffstat:

 raid0.c |   21 +++++----------------
 raid1.c |   14 +++-----------
 2 files changed, 8 insertions(+), 27 deletions(-)

Thanks for the patchset. I applied it to the md-6.11 branch.

I was planning to respin this to also handle failures in the takeover path.

maintainer
#8 lore raw

Hi Christoph,

Quoted 11 lines · Christoph Hellwig
Quoted 9 lines · Song Liu
Quoted 8 lines

raid10 and raid5 also free conf on failure in their ->run handle, but set mddev->private to NULL which prevents the double free as well (but requires more code)

Diffstat:

 raid0.c |   21 +++++----------------
 raid1.c |   14 +++-----------
 2 files changed, 8 insertions(+), 27 deletions(-)
Quoted 4 lines
Quoted 1 line

Thanks for the patchset. I applied it to the md-6.11 branch.

I was planning to respin this to also handle failures in the takeover path.

How about we do that in a follow up patch?

Thanks for fixing these issues!

Song