Skip to contentSkip to conversation

Fix nested permission update

PATCH In review Vladimir Sementsov-Ogievskiy 3 participants 10 messages Oct 31, 2020 – Nov 6, 2020

lore
Share & tools
lore
#0 lore raw

cover Fix nested permission update

Hi! With help of some assertions (patch 2) I've found that bdrv_drop_intermediate() do nested permission update which in my opinion may lead to unpredictable behavior.

Vladimir Sementsov-Ogievskiy (2):
block: make bdrv_drop_intermediate() less wrong
block: assert that permission commit sets same permissions

 block.c | 47 ++++++++++++++++++++++++-----------------------
 1 file changed, 24 insertions(+), 23 deletions(-)
Signature · 1 line
2.21.3
#1 lore raw

PATCH 1/2 block: make bdrv_drop_intermediate() less wrong

First, permission update loop tries to do iterations transactionally, but the whole update is not transactional: nobody roll-back successful loop iterations when some iteration fails.

Second, in the iteration we have nested permission update:
c->klass->update_filename may point to bdrv_child_cb_update_filename() which calls bdrv_backing_update_filename(), which may do node reopen to RW.

Permission update system is not prepared to nested updates, at least it has intermediate permission-update state stored in BdrvChild
structures: has_backup_perm, backup_perm and backup_shared_perm.

So, let's first do bdrv_replace_node() (which is more transactional than open-coded update in bdrv_drop_intermediate()) and then call update_filename() in separate. We still do not rollback changes in case of update_filename() failure but it's not much worse than pre-patch behavior.

---
 block.c | 36 +++++++++++++++---------------------
 1 file changed, 15 insertions(+), 21 deletions(-)
block.c +15 -21
index 9538af4884..bd9f4e534b 100644
@@ -4958,36 +4958,30 @@ int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
         backing_file_str = base->filename;
     }
 
-    QLIST_FOREACH_SAFE(c, &top->parents, next_parent, next) {
-        /* Check whether we are allowed to switch c from top to base */
-        GSList *ignore_children = g_slist_prepend(NULL, c);
-        ret = bdrv_check_update_perm(base, NULL, c->perm, c->shared_perm,
-                                     ignore_children, NULL, &local_err);
-        g_slist_free(ignore_children);
-        if (ret < 0) {
-            error_report_err(local_err);
-            goto exit;
-        }
+    bdrv_replace_node(top, base, &local_err);
+    if (local_err) {
+        error_report_err(local_err);
+        goto exit;
+    }
 
-        /* If so, update the backing file path in the image file */
+    QLIST_FOREACH_SAFE(c, &base->parents, next_parent, next) {
         if (c->klass->update_filename) {
             ret = c->klass->update_filename(c, base, backing_file_str,
                                             &local_err);
             if (ret < 0) {
-                bdrv_abort_perm_update(base);
+                /*
+                 * TODO: Actually, we want to rollback all previous iterations
+                 * of this loop, and (which is almost impossible) previous
+                 * bdrv_replace_node()...
+                 *
+                 * Note, that c->klass->update_filename may lead to permission
+                 * update, so it's a bad idea to call it inside permission
+                 * update transaction of bdrv_replace_node.
+                 */
                 error_report_err(local_err);
                 goto exit;
             }
         }
-
-        /*
-         * Do the actual switch in the in-memory graph.
-         * Completes bdrv_check_update_perm() transaction internally.
-         * c->frozen is false, we have checked that above.
-         */
-        bdrv_ref(base);
-        bdrv_replace_child(c, base);
-        bdrv_unref(top);
     }
 
     if (update_inherits_from) {
Signature · 1 line
2.21.3
#7 lore raw

PATCH 2/2 block: assert that permission commit sets same permissions

In reply to Vladimir Sementsov-Ogievskiy

On permission update commit we must set same permissions as on _check_. Let's add assertions. Next step may be to drop permission parameters from _set_.

Note that prior to previous commit, fixing bdrv_drop_intermediate(), new assertion in bdrv_child_set_perm() crashes on iotests 30 and 40.

---
 block.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
block.c +9 -2
index bd9f4e534b..0f4da59a6c 100644
@@ -2105,9 +2105,10 @@ static void bdrv_abort_perm_update(BlockDriverState *bs)
     }
 }
 
-static void bdrv_set_perm(BlockDriverState *bs, uint64_t cumulative_perms,
-                          uint64_t cumulative_shared_perms)
+static void bdrv_set_perm(BlockDriverState *bs, uint64_t _cumulative_perms,
+                          uint64_t _cumulative_shared_perms)
 {
+    uint64_t cumulative_perms, cumulative_shared_perms;
     BlockDriver *drv = bs->drv;
     BdrvChild *c;
 
@@ -2115,6 +2116,10 @@ static void bdrv_set_perm(BlockDriverState *bs, uint64_t cumulative_perms,
         return;
     }
 
+    bdrv_get_cumulative_perm(bs, &cumulative_perms, &cumulative_shared_perms);
+    assert(_cumulative_perms == cumulative_perms);
+    assert(_cumulative_shared_perms == cumulative_shared_perms);
+
     /* Update this node */
     if (drv->bdrv_set_perm) {
         drv->bdrv_set_perm(bs, cumulative_perms, cumulative_shared_perms);
@@ -2301,6 +2306,8 @@ static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)
 
     c->has_backup_perm = false;
 
+    assert(c->perm == perm);
+    assert(c->shared_perm == shared);
     c->perm = perm;
     c->shared_perm = shared;
 
Signature · 1 line
2.21.3
#8 lore raw

Re: [PATCH 2/2] block: assert that permission commit sets same permissions

Quoted 42 lines · Vladimir Sementsov-Ogievskiy

On permission update commit we must set same permissions as on _check_. Let's add assertions. Next step may be to drop permission parameters from _set_.

Note that prior to previous commit, fixing bdrv_drop_intermediate(), new assertion in bdrv_child_set_perm() crashes on iotests 30 and 40.

---
  block.c | 11 +++++++++--
  1 file changed, 9 insertions(+), 2 deletions(-)
block.c +7 -2
index bd9f4e534b..0f4da59a6c 100644
@@ -2105,9 +2105,10 @@ static void bdrv_abort_perm_update(BlockDriverState *bs)
      }
  }
 
-static void bdrv_set_perm(BlockDriverState *bs, uint64_t cumulative_perms,
-                          uint64_t cumulative_shared_perms)
+static void bdrv_set_perm(BlockDriverState *bs, uint64_t _cumulative_perms,
+                          uint64_t _cumulative_shared_perms)
  {
+    uint64_t cumulative_perms, cumulative_shared_perms;
      BlockDriver *drv = bs->drv;
      BdrvChild *c;
 
@@ -2115,6 +2116,10 @@ static void bdrv_set_perm(BlockDriverState *bs, uint64_t cumulative_perms,
          return;
      }
 
+    bdrv_get_cumulative_perm(bs, &cumulative_perms, &cumulative_shared_perms);
+    assert(_cumulative_perms == cumulative_perms);
+    assert(_cumulative_shared_perms == cumulative_shared_perms);
+
      /* Update this node */
      if (drv->bdrv_set_perm) {
          drv->bdrv_set_perm(bs, cumulative_perms, cumulative_shared_perms);

block.c static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared) · line 2313

Quoted 8 lines
block.c +2 -0
@@ -2301,6 +2306,8 @@ static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)
 
      c->has_backup_perm = false;
 
+    assert(c->perm == perm);
+    assert(c->shared_perm == shared);
      c->perm = perm;
      c->shared_perm = shared;
 

squash:

@@ -2308,8 +2308,6 @@ static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)

     assert(c->perm == perm);
     assert(c->shared_perm == shared);

- c->perm = perm;
- c->shared_perm = shared;

     bdrv_get_cumulative_perm(c->bs, &cumulative_perms,
                              &cumulative_shared_perms);
Signature · 2 lines
Best regards,
Vladimir
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] block: make bdrv_drop_intermediate() less wrong

In reply to Vladimir Sementsov-Ogievskiy
Quoted 3 lines · Vladimir Sementsov-Ogievskiy

First, permission update loop tries to do iterations transactionally, but the whole update is not transactional: nobody roll-back successful loop iterations when some iteration fails.

Indeed. Another thing that should be noted:

bdrv_check_update_perm() doc:

Quoted 2 lines

Needs to be followed by a call to either bdrv_set_perm() or bdrv_abort_perm_update().

And besides rolling back on error, we don’t commit on success either.

Quoted 15 lines

Second, in the iteration we have nested permission update:
c->klass->update_filename may point to bdrv_child_cb_update_filename() which calls bdrv_backing_update_filename(), which may do node reopen to RW.

Permission update system is not prepared to nested updates, at least it has intermediate permission-update state stored in BdrvChild
structures: has_backup_perm, backup_perm and backup_shared_perm.

So, let's first do bdrv_replace_node() (which is more transactional than open-coded update in bdrv_drop_intermediate()) and then call update_filename() in separate. We still do not rollback changes in case of update_filename() failure but it's not much worse than pre-patch behavior.

Quoted 4 lines
---
  block.c | 36 +++++++++++++++---------------------
  1 file changed, 15 insertions(+), 21 deletions(-)

And it’s also much simpler and clearer now.

#3 lore raw

Re: [PATCH 1/2] block: make bdrv_drop_intermediate() less wrong

Quoted 4 lines · Max Reitz
Quoted 3 lines · Vladimir Sementsov-Ogievskiy

First, permission update loop tries to do iterations transactionally, but the whole update is not transactional: nobody roll-back successful loop iterations when some iteration fails.

[...]

Quoted 1 line

And besides rolling back on error, we don’t commit on success either.

(Oh, we do, through bdrv_replace_child(). Well.)

Max

Reviewed-by on v1 2/2
#9 lore raw

Re: [PATCH 2/2] block: assert that permission commit sets same permissions

In reply to Vladimir Sementsov-Ogievskiy
Quoted 16 lines · Vladimir Sementsov-Ogievskiy

On permission update commit we must set same permissions as on _check_. Let's add assertions. Next step may be to drop permission parameters from _set_.

Note that prior to previous commit, fixing bdrv_drop_intermediate(), new assertion in bdrv_child_set_perm() crashes on iotests 30 and 40.

---
  block.c | 11 +++++++++--
  1 file changed, 9 insertions(+), 2 deletions(-)
block.c +0 -0
index bd9f4e534b..0f4da59a6c 100644

[...]

block.c static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared) · line 2312

Quoted 8 lines
block.c +2 -0
@@ -2301,6 +2306,8 @@ static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)
 
      c->has_backup_perm = false;
 
+    assert(c->perm == perm);
+    assert(c->shared_perm == shared);
      c->perm = perm;
      c->shared_perm = shared;

Then we can drop the assignments, no?

(And, as you write, in the future potentially drop the parameters.)

Anyway:

#4 lore raw

Re: [PATCH 1/2] block: make bdrv_drop_intermediate() less wrong

In reply to Vladimir Sementsov-Ogievskiy
Quoted 9 lines · Vladimir Sementsov-Ogievskiy
block.c +0 -5
@@ -4958,36 +4958,30 @@ int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
         backing_file_str = base->filename;
     }
 
-    QLIST_FOREACH_SAFE(c, &top->parents, next_parent, next) {
-        /* Check whether we are allowed to switch c from top to base */
-        GSList *ignore_children = g_slist_prepend(NULL, c);
-        ret = bdrv_check_update_perm(base, NULL, c->perm, c->shared_perm,
-                                     ignore_children, NULL, &local_err);

block.c int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base, · line 4965

Quoted 10 lines
block.c +5 -5
-        g_slist_free(ignore_children);
-        if (ret < 0) {
-            error_report_err(local_err);
-            goto exit;
-        }
+    bdrv_replace_node(top, base, &local_err);
+    if (local_err) {
+        error_report_err(local_err);
+        goto exit;
+    }

At the beginning of this function there's a check for c->frozen. I think you can remove it safely because you also have it in bdrv_replace_node()

Berto

#5 lore raw

Re: [PATCH 1/2] block: make bdrv_drop_intermediate() less wrong

In reply to Vladimir Sementsov-Ogievskiy
Quoted 1 line · Vladimir Sementsov-Ogievskiy

- QLIST_FOREACH_SAFE(c, &top->parents, next_parent, next) {

/* ... */

Quoted 1 line

+ QLIST_FOREACH_SAFE(c, &base->parents, next_parent, next) {

I also wonder, is top->parents and base->parents guaranteed to be the same list in this case?

If not you could store the list of top->parents before calling bdrv_replace_node() and use it afterwards.

    QLIST_FOREACH(c, &top->parents, next_parent) {
        parents = g_slist_prepend(parents, c);
    }

Berto

#6 lore raw

Re: [PATCH 1/2] block: make bdrv_drop_intermediate() less wrong

Quoted 8 lines · Alberto Garcia
Quoted 1 line · Vladimir Sementsov-Ogievskiy

- QLIST_FOREACH_SAFE(c, &top->parents, next_parent, next) {

/* ... */

Quoted 1 line

+ QLIST_FOREACH_SAFE(c, &base->parents, next_parent, next) {

I also wonder, is top->parents and base->parents guaranteed to be the same list in this case?

Quoted 9 lines

If not you could store the list of top->parents before calling bdrv_replace_node() and use it afterwards.

     QLIST_FOREACH(c, &top->parents, next_parent) {
         parents = g_slist_prepend(parents, c);
     }

Berto

Hmm... We should not touch other parents of base, which it had prior to bdrv_replace_node().

On the other hand, it should be safe to call update_filename for not-updated parents..

And we should keep in mind that bdrv_replace_node may replace not all children. Still, in this case we must replace all of them. Seems, we need additional argument for bdrv_replace_node() to fail, if it want's to skip some children replacement.

So, I'm going to resend to add this parameter to bdrv_replace_node() and will use your suggestion to be stricter on what we update, thanks!

Signature · 2 lines
Best regards,
Vladimir