Skip to contentSkip to conversation

blkdebug: Don't leak bs->file on failure

PATCH Superseded Kevin Wolf 2 participants 4 messages Feb 8, 2014 Block layer

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Kevin Wolf

author
#0 lore raw

PATCH 1/1 blkdebug: Don't leak bs->file on failure

---
 block/blkdebug.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
block/blkdebug.c +6 -3
index 56c4cd0..519b483 100644
@@ -414,7 +414,7 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
                           flags, true, false, &local_err);
     if (ret < 0) {
         error_propagate(errp, local_err);
-        goto fail;
+        goto fail_unref;
     }
 
     /* Set request alignment */
@@ -424,10 +424,13 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
     } else {
         error_setg(errp, "Invalid alignment");
         ret = -EINVAL;
-        goto fail;
+        goto fail_unref;
     }
 
-    ret = 0;
+    return 0;
+
+fail_unref:
+    bdrv_unref(bs->file);
 fail:
     qemu_opts_del(opts);
     return ret;
Signature · 1 line
1.8.1.4
#1 lore raw

comments below

Quoted 10 lines · Kevin Wolf
---
 block/blkdebug.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
block/blkdebug.c +0 -0
index 56c4cd0..519b483 100644

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 418

Quoted 6 lines
block/blkdebug.c +1 -1
@@ -414,7 +414,7 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
                           flags, true, false, &local_err);
     if (ret < 0) {
         error_propagate(errp, local_err);
-        goto fail;
+        goto fail_unref;
     }

This block appears to run when the bdrv_open_image(&bs->file, ...) call fails. If we can't open the backing file, are you sure we need to unref it under fail_unref? I thought that this block needed no update (one only needs to extend the error handling when allocation of a new resource is successful), but I may have been wrong of course. (Generally in qemu it's always anyone's best guess who is responsible for releasing stuff on the error path, caller or callee.)

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 428

Quoted 9 lines
block/blkdebug.c +1 -1
     /* Set request alignment */
@@ -424,10 +424,13 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
     } else {
         error_setg(errp, "Invalid alignment");
         ret = -EINVAL;
-        goto fail;
+        goto fail_unref;
     }

I do agree with this.

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 436

Quoted 10 lines
block/blkdebug.c +4 -1
-    ret = 0;
+    return 0;
+
+fail_unref:
+    bdrv_unref(bs->file);
 fail:
     qemu_opts_del(opts);
     return ret;
 

I believe this change causes us to leak "opts" on success.

... Unless we should have hung on to it to begin with, of course, in which case this is a silent bugfix. But I don't believe so; I think we only need "opts" temporarily, to absorb options from the QDict, and for ease of parsing into local variables.

How about the attached patch instead? Releasing stuff in specific error handler blocks quickly becomes intractable if there are many resources to allocate in succession, but in this case I think we can easily get away with it.

Thanks,
Laszlo

block/blkdebug.c +6 -5
index 56c4cd0..1a8ccc4 100644
@@ -396,14 +396,14 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
     if (error_is_set(&local_err)) {
         error_propagate(errp, local_err);
         ret = -EINVAL;
-        goto fail;
+        goto done;
     }
 
     /* Read rules from config file or command line options */
     config = qemu_opt_get(opts, "config");
     ret = read_config(s, config, options, errp);
     if (ret) {
-        goto fail;
+        goto done;
     }
 
     /* Set initial state */
@@ -414,7 +414,7 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
                           flags, true, false, &local_err);
     if (ret < 0) {
         error_propagate(errp, local_err);
-        goto fail;
+        goto done;
     }
 
     /* Set request alignment */
@@ -423,12 +423,13 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
         bs->request_alignment = align;
     } else {
         error_setg(errp, "Invalid alignment");
+        bdrv_unref(bs->file);
         ret = -EINVAL;
-        goto fail;
+        goto done;
     }
 
     ret = 0;
-fail:
+done:
     qemu_opts_del(opts);
     return ret;
 }
author
#2 lore raw

Am 08.02.2014 um 10:46 hat Laszlo Ersek geschrieben:

Quoted 53 lines

comments below

Quoted 10 lines · Kevin Wolf
---
 block/blkdebug.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
block/blkdebug.c +0 -0
index 56c4cd0..519b483 100644

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 418

Quoted 6 lines
block/blkdebug.c +1 -1
@@ -414,7 +414,7 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
                           flags, true, false, &local_err);
     if (ret < 0) {
         error_propagate(errp, local_err);
-        goto fail;
+        goto fail_unref;
     }

This block appears to run when the bdrv_open_image(&bs->file, ...) call fails. If we can't open the backing file, are you sure we need to unref it under fail_unref? I thought that this block needed no update (one only needs to extend the error handling when allocation of a new resource is successful), but I may have been wrong of course. (Generally in qemu it's always anyone's best guess who is responsible for releasing stuff on the error path, caller or callee.)

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 428

Quoted 9 lines
block/blkdebug.c +1 -1
     /* Set request alignment */
@@ -424,10 +424,13 @@ static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
     } else {
         error_setg(errp, "Invalid alignment");
         ret = -EINVAL;
-        goto fail;
+        goto fail_unref;
     }

I do agree with this.

block/blkdebug.c static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags, · line 436

Quoted 10 lines
block/blkdebug.c +4 -1
-    ret = 0;
+    return 0;
+
+fail_unref:
+    bdrv_unref(bs->file);
 fail:
     qemu_opts_del(opts);
     return ret;
 

I believe this change causes us to leak "opts" on success.

Quoted 4 lines

... Unless we should have hung on to it to begin with, of course, in which case this is a silent bugfix. But I don't believe so; I think we only need "opts" temporarily, to absorb options from the QDict, and for ease of parsing into local variables.

As always, you're right with all you're saying. Sorry for being sloppy.

Quoted 4 lines

How about the attached patch instead? Releasing stuff in specific error handler blocks quickly becomes intractable if there are many resources to allocate in succession, but in this case I think we can easily get away with it.

I'd prefer to avoid this because it's too easy to miss it when you add new code to the function. Let me send a v2 that keeps fail_unref.

Kevin

#3 lore raw
Quoted 9 lines · Kevin Wolf

Am 08.02.2014 um 10:46 hat Laszlo Ersek geschrieben:

Quoted 4 lines

How about the attached patch instead? Releasing stuff in specific error handler blocks quickly becomes intractable if there are many resources to allocate in succession, but in this case I think we can easily get away with it.

I'd prefer to avoid this because it's too easy to miss it when you add new code to the function.

Oh, right. If you actually care about the future, this is a good point :)

Thanks!
Laszlo