Skip to contentSkip to conversation

qcow2: don't leave partially initialized file on image creation

PATCH Superseded Maxim Levitsky 2 participants 11 messages Dec 9, 2020 – Dec 10, 2020 Block layer

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v5 cover qcow2: don't leave partially initialized file on image creation

Use the bdrv_co_delete_file interface to delete the underlying file if qcow2 initialization fails (e.g due to bad encryption secret)

This makes the qcow2 driver behave the same way as the luks driver behaves.

V3: addressed review feedback and reworked commit messages

V4: got rid of code duplication by adding bdrv_co_delete_file_noerr and made the qcow2 driver use this function to delete both the main and the data file.

V5: addresssed review feedback on reworked version.

Best regards,
Maxim Levitsky

Maxim Levitsky (4):
crypto: luks: Fix tiny memory leak
block: add bdrv_co_delete_file_noerr
crypto: luks: use bdrv_co_delete_file_noerr
block: qcow2: remove the created file on initialization error

 block.c               | 24 ++++++++++++++++++++++++
 block/crypto.c        | 13 ++-----------
 block/qcow2.c         |  6 ++++--
 include/block/block.h |  1 +
 4 files changed, 31 insertions(+), 13 deletions(-)
Signature · 1 line
2.26.2
#1 lore raw

PATCH v5 1/4 crypto: luks: Fix tiny memory leak

When the underlying block device doesn't support the bdrv_co_delete_file interface, an 'Error' object was leaked.

---
 block/crypto.c | 2 ++
 1 file changed, 2 insertions(+)
block/crypto.c +2 -0
index aef5a5721a..b3a5275132 100644
@@ -735,6 +735,8 @@ fail:
          */
         if ((r_del < 0) && (r_del != -ENOTSUP)) {
             error_report_err(local_delete_err);
+        } else {
+            error_free(local_delete_err);
         }
     }
 
Signature · 1 line
2.26.2
Reviewed-by on v5 1/4
#2 lore raw

Re: [PATCH v5 1/4] crypto: luks: Fix tiny memory leak

Quoted 5 lines · Maxim Levitsky

When the underlying block device doesn't support the bdrv_co_delete_file interface, an 'Error' object was leaked.

Quoted 18 lines
---
  block/crypto.c | 2 ++
  1 file changed, 2 insertions(+)
block/crypto.c +2 -0
index aef5a5721a..b3a5275132 100644
@@ -735,6 +735,8 @@ fail:
           */
          if ((r_del < 0) && (r_del != -ENOTSUP)) {
              error_report_err(local_delete_err);
+        } else {
+            error_free(local_delete_err);
          }
      }
 
Signature · 2 lines
Best regards,
Vladimir
#3 lore raw

PATCH v5 2/4 block: add bdrv_co_delete_file_noerr

In reply to Maxim Levitsky

This function wraps bdrv_co_delete_file for the common case of removing a file, which was just created by format driver, on an error condition.

It hides the -ENOTSUPP error, and reports all other errors otherwise.

---
 block.c               | 24 ++++++++++++++++++++++++
 include/block/block.h |  1 +
 2 files changed, 25 insertions(+)
block.c +24 -0
index f1cedac362..5d35ba2fb8 100644
@@ -704,6 +704,30 @@ int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp)
     return ret;
 }
 
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs)
+{
+    Error *local_err = NULL;
+    int ret;
+
+    if (!bs) {
+        return;
+    }
+
+    ret = bdrv_co_delete_file(bs, &local_err);
+    /*
+     * ENOTSUP will happen if the block driver doesn't support
+     * the 'bdrv_co_delete_file' interface. This is a predictable
+     * scenario and shouldn't be reported back to the user.
+     */
+    if (ret == -ENOTSUP) {
+        error_free(local_err);
+    } else if (ret < 0) {
+        error_report_err(local_err);
+    }
+}
+
+
+
 /**
  * Try to get @bs's logical and physical block size.
  * On success, store them in @bsz struct and return 0.
include/block/block.h +1 -0
index c9d7c58765..af03022723 100644
@@ -428,6 +428,7 @@ int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base,
                               Error **errp);
 void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base);
 int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp);
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs);
 
 
 typedef struct BdrvCheckResult {
Signature · 1 line
2.26.2
Reviewed-by on v5 2/4
#4 lore raw

Re: [PATCH v5 2/4] block: add bdrv_co_delete_file_noerr

Quoted 4 lines · Maxim Levitsky

This function wraps bdrv_co_delete_file for the common case of removing a file, which was just created by format driver, on an error condition.

It hides the -ENOTSUPP error, and reports all other errors otherwise.

I've looked at original commit added this logic, and didn't find exactly, why should we hide it..

Quoted 21 lines
---
  block.c               | 24 ++++++++++++++++++++++++
  include/block/block.h |  1 +
  2 files changed, 25 insertions(+)
block.c +5 -0
index f1cedac362..5d35ba2fb8 100644
@@ -704,6 +704,30 @@ int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp)
      return ret;
  }
 
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs)
+{
+    Error *local_err = NULL;
+    int ret;
+

block.c int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp) · line 721

Quoted 10 lines
block.c +10 -0
+    if (!bs) {
+        return;
+    }
+
+    ret = bdrv_co_delete_file(bs, &local_err);
+    /*
+     * ENOTSUP will happen if the block driver doesn't support
+     * the 'bdrv_co_delete_file' interface. This is a predictable
+     * scenario and shouldn't be reported back to the user.
+     */

It's not predictable for user: user doesn't know internal processes and interfaces.

The problem: we've left extra file on failure scenario and can't remove it. We want to warn the user that there is a wrong file left. Why not to warn, when the removement is unsupported internally by the driver?

I think, it's better to report it in any case.

Another reason: less code and logic on failure case. Optimizing failure scenarios in different manner is seldom a good thing to do.

And finally: why to report the error at all? If we have errp parameter, it's better to add the info to it using error_append.. something like error_append(errp, " (also failed to remove unfinished file %s: %s)", file_name, error_get_pretty(local_err))

Probably I'm making mountains out of molehills. It should work, so take my r-b anyway:

Side note: I'd not split patches 02 and 03: this way it would be simpler to see the code movement.

block.c int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp) · line 730

Quoted 9 lines
block.c +9 -0
+    if (ret == -ENOTSUP) {
+        error_free(local_err);
+    } else if (ret < 0) {
+        error_report_err(local_err);
+    }
+}
+
+
+

three empty lines..

Quoted 16 lines
block.c +0 -0
  /**
   * Try to get @bs's logical and physical block size.
   * On success, store them in @bsz struct and return 0.
include/block/block.h +1 -0
index c9d7c58765..af03022723 100644
@@ -428,6 +428,7 @@ int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base,
                                Error **errp);
  void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base);
  int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp);
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs);
 
 
  typedef struct BdrvCheckResult {
Signature · 2 lines
Best regards,
Vladimir
#5 lore raw

Re: [PATCH v5 2/4] block: add bdrv_co_delete_file_noerr

Quoted 41 lines · Vladimir Sementsov-Ogievskiy
Quoted 4 lines · Maxim Levitsky

This function wraps bdrv_co_delete_file for the common case of removing a file, which was just created by format driver, on an error condition.

It hides the -ENOTSUPP error, and reports all other errors otherwise.

I've looked at original commit added this logic, and didn't find exactly, why should we hide it..

Quoted 20 lines
---
  block.c               | 24 ++++++++++++++++++++++++
  include/block/block.h |  1 +
  2 files changed, 25 insertions(+)
block.c +5 -0
index f1cedac362..5d35ba2fb8 100644
@@ -704,6 +704,30 @@ int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp)
      return ret;
  }
 
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs)
+{
+    Error *local_err = NULL;
+    int ret;
+

block.c int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp) · line 721

Quoted 10 lines
block.c +10 -0
+    if (!bs) {
+        return;
+    }
+
+    ret = bdrv_co_delete_file(bs, &local_err);
+    /*
+     * ENOTSUP will happen if the block driver doesn't support
+     * the 'bdrv_co_delete_file' interface. This is a predictable
+     * scenario and shouldn't be reported back to the user.
+     */

It's not predictable for user: user doesn't know internal processes and interfaces.

Quoted 3 lines

The problem: we've left extra file on failure scenario and can't remove it. We want to warn the user that there is a wrong file left. Why not to warn, when the removement is unsupported internally by the driver?

I think, it's better to report it in any case.

This is a good point, but it is not something I changed. I prefer this to be changed in a separate patch.

Quoted 4 lines

Another reason: less code and logic on failure case. Optimizing failure scenarios in different manner is seldom a good thing to do.

And finally: why to report the error at all? If we have errp parameter, it's better to add the info to it using error_append.. something like error_append(errp, " (also failed to remove unfinished file %s: %s)", file_name, error_get_pretty(local_err))

The idea behind this was to reduce code duplication in the callers of this function, and this is why this function doesn't even have the 'errp' parameter. I'll think about it.

Quoted 9 lines

Probably I'm making mountains out of molehills. It should work, so take my r-b anyway:

Side note: I'd not split patches 02 and 03: this way it would be simpler to see the code movement.

Quoted 11 lines

block.c int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp) · line 730

Quoted 9 lines
block.c +9 -0
+    if (ret == -ENOTSUP) {
+        error_free(local_err);
+    } else if (ret < 0) {
+        error_report_err(local_err);
+    }
+}
+
+
+

three empty lines..

Will fix.

Quoted 15 lines
Quoted 15 lines
block.c +0 -0
  /**
   * Try to get @bs's logical and physical block size.
   * On success, store them in @bsz struct and return 0.
include/block/block.h +1 -0
index c9d7c58765..af03022723 100644
@@ -428,6 +428,7 @@ int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base,
                                Error **errp);
  void bdrv_unfreeze_backing_chain(BlockDriverState *bs, BlockDriverState *base);
  int coroutine_fn bdrv_co_delete_file(BlockDriverState *bs, Error **errp);
+void coroutine_fn bdrv_co_delete_file_noerr(BlockDriverState *bs);
 
 
Quoted 2 lines

include/block/block.h int bdrv_freeze_backing_chain(BlockDriverState *bs, BlockDriverState *base, · line 434

Quoted 2 lines
include/block/block.h +0 -0
  typedef struct BdrvCheckResult {

Thanks a lot for the review!

Best regards,
Maxim Levitsky

Quoted 2 lines
#6 lore raw

PATCH v5 3/4 crypto: luks: use bdrv_co_delete_file_noerr

In reply to Maxim Levitsky

This refactoring is now possible thanks to this function.

---
 block/crypto.c | 15 ++-------------
 1 file changed, 2 insertions(+), 13 deletions(-)
block/crypto.c +2 -13
index b3a5275132..1d30fde38e 100644
@@ -725,19 +725,8 @@ fail:
      * If an error occurred, delete 'filename'. Even if the file existed
      * beforehand, it has been truncated and corrupted in the process.
      */
-    if (ret && bs) {
-        Error *local_delete_err = NULL;
-        int r_del = bdrv_co_delete_file(bs, &local_delete_err);
-        /*
-         * ENOTSUP will happen if the block driver doesn't support
-         * the 'bdrv_co_delete_file' interface. This is a predictable
-         * scenario and shouldn't be reported back to the user.
-         */
-        if ((r_del < 0) && (r_del != -ENOTSUP)) {
-            error_report_err(local_delete_err);
-        } else {
-            error_free(local_delete_err);
-        }
+    if (ret) {
+        bdrv_co_delete_file_noerr(bs);
     }
 
     bdrv_unref(bs);
Signature · 1 line
2.26.2
#8 lore raw

PATCH v5 4/4 block: qcow2: remove the created file on initialization error

In reply to Maxim Levitsky

If the qcow initialization fails, we should remove the file if it was already created, to avoid leaving stale files around.

We already do this for luks raw images.

---
 block/qcow2.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
block/qcow2.c +4 -2
index 3a90ef2786..68c9182f92 100644
@@ -3847,12 +3847,14 @@ static int coroutine_fn qcow2_co_create_opts(BlockDriver *drv,
 
     /* Create the qcow2 image (format layer) */
     ret = qcow2_co_create(create_options, errp);
+finish:
     if (ret < 0) {
-        goto finish;
+        bdrv_co_delete_file_noerr(bs);
+        bdrv_co_delete_file_noerr(data_bs);
     }
 
     ret = 0;
-finish:
+
     qobject_unref(qdict);
     bdrv_unref(bs);
     bdrv_unref(data_bs);
Signature · 1 line
2.26.2
#9 lore raw

Re: [PATCH v5 4/4] block: qcow2: remove the created file on initialization error

Quoted 17 lines · Maxim Levitsky

If the qcow initialization fails, we should remove the file if it was already created, to avoid leaving stale files around.

We already do this for luks raw images.

---
  block/qcow2.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
block/qcow2.c +0 -0
index 3a90ef2786..68c9182f92 100644
@@ -3847,12 +3847,14 @@ static int coroutine_fn qcow2_co_create_opts(BlockDriver *drv,
 

block/qcow2.c static int coroutine_fn qcow2_co_create_opts(BlockDriver *drv, · line 3856

Quoted 10 lines
block/qcow2.c +3 -1
      /* Create the qcow2 image (format layer) */
      ret = qcow2_co_create(create_options, errp);
+finish:
      if (ret < 0) {
-        goto finish;
+        bdrv_co_delete_file_noerr(bs);
+        bdrv_co_delete_file_noerr(data_bs);
      }
 
      ret = 0;

Ooops :) After this patch the function always returns zero :)

Quoted 6 lines
block/qcow2.c +1 -1
-finish:
+
      qobject_unref(qdict);
      bdrv_unref(bs);
      bdrv_unref(data_bs);
Signature · 2 lines
Best regards,
Vladimir
#10 lore raw

Re: [PATCH v5 4/4] block: qcow2: remove the created file on initialization error

Quoted 27 lines · Vladimir Sementsov-Ogievskiy
Quoted 26 lines · Maxim Levitsky

If the qcow initialization fails, we should remove the file if it was already created, to avoid leaving stale files around.

We already do this for luks raw images.

---
  block/qcow2.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
block/qcow2.c +3 -1
index 3a90ef2786..68c9182f92 100644
@@ -3847,12 +3847,14 @@ static int coroutine_fn qcow2_co_create_opts(BlockDriver *drv,
 
      /* Create the qcow2 image (format layer) */
      ret = qcow2_co_create(create_options, errp);
+finish:
      if (ret < 0) {
-        goto finish;
+        bdrv_co_delete_file_noerr(bs);
+        bdrv_co_delete_file_noerr(data_bs);
      }
 
Quoted 3 lines

block/qcow2.c static int coroutine_fn qcow2_co_create_opts(BlockDriver *drv, · line 3856

Quoted 1 line
block/qcow2.c +0 -0
      ret = 0;

Ooops :) After this patch the function always returns zero :)

Indeed :-(

These small changes done in the last minute are the most dangerous.

Best regards,
Maxim Levitsky

Quoted 9 lines
Quoted 7 lines
block/qcow2.c +1 -1
-finish:
+
      qobject_unref(qdict);
      bdrv_unref(bs);
      bdrv_unref(data_bs);