diff mbox series

[1/1] block/file-posix.c: fix not reopened lock file descriptor

Message ID 20180322172053.7343-1-dionbosschieter@gmail.com
State New
Headers show
Series [1/1] block/file-posix.c: fix not reopened lock file descriptor | expand

Commit Message

Dion Bosschieter March 22, 2018, 5:20 p.m. UTC
In commit 244a5668106297378391b768e7288eb157616f64 another
file descriptor to BDRVRawState is added. When we try to issue the
reopen command only s->fd is reopened; lock_fd could still hold an old
file descriptor "possibly" pointing to another file.

- change raw_reopen_prepare so it checks use_lock from BDRVRawState and
tries to reopen lock_fd accordingly
- change raw_reopen_commit so it closes the old lock_fd on use_lock

Signed-off-by: Dion Bosschieter <dionbosschieter@gmail.com>
---
 block/file-posix.c | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

Comments

Dion Bosschieter March 22, 2018, 6:08 p.m. UTC | #1
Yeah I have a use case, before a last sync on a storage migration we suspend a VM -> send the last diffs -> mount the new storage server and after that we change a symlink -> call reopen -> check if all file descriptors are changed before resuming the VM.

Dion

> Op 22 mrt. 2018 om 18:39 heeft Kevin Wolf <kwolf@redhat.com> het volgende geschreven:
> 
> [ Cc: qemu-block ]
> 
> Am 22.03.2018 um 18:20 hat Dion Bosschieter geschrieben:
>> In commit 244a5668106297378391b768e7288eb157616f64 another
>> file descriptor to BDRVRawState is added. When we try to issue the
>> reopen command only s->fd is reopened; lock_fd could still hold an old
>> file descriptor "possibly" pointing to another file.
>> 
>> - change raw_reopen_prepare so it checks use_lock from BDRVRawState and
>> tries to reopen lock_fd accordingly
>> - change raw_reopen_commit so it closes the old lock_fd on use_lock
>> 
>> Signed-off-by: Dion Bosschieter <dionbosschieter@gmail.com>
> 
> bdrv_reopen() is not meant for opening a different file, it is meant to
> change the flags and options of the same file. Do you have a use case
> where you would actually need to switch to a different file?
> 
> As far as I know, lock_fd was specifically introduced _because_ it stays
> the same across reopen, so we don't need a racy release/reacquire pair.
> Fam (CCed) should know more.
> 
> In any case, doesn't your patch drop all the locks without reacquiring
> them on the new lock_fd?
> 
> Kevin
> 
>> block/file-posix.c | 25 +++++++++++++++++++++++++
>> 1 file changed, 25 insertions(+)
>> 
>> diff --git a/block/file-posix.c b/block/file-posix.c
>> index d7fb772c14..16d83fc49e 100644
>> --- a/block/file-posix.c
>> +++ b/block/file-posix.c
>> @@ -167,6 +167,7 @@ typedef struct BDRVRawState {
>> 
>> typedef struct BDRVRawReopenState {
>>     int fd;
>> +    int lock_fd;
>>     int open_flags;
>> } BDRVRawReopenState;
>> 
>> @@ -795,6 +796,7 @@ static int raw_reopen_prepare(BDRVReopenState *state,
>>     raw_parse_flags(state->flags, &rs->open_flags);
>> 
>>     rs->fd = -1;
>> +    rs->lock_fd = -1;
>> 
>>     int fcntl_flags = O_APPEND | O_NONBLOCK;
>> #ifdef O_NOATIME
>> @@ -820,6 +822,17 @@ static int raw_reopen_prepare(BDRVReopenState *state,
>>                 rs->fd = -1;
>>             }
>>         }
>> +
>> +        if (s->use_lock) {
>> +            rs->lock_fd = qemu_dup(s->lock_fd);
>> +            if (rs->lock_fd >= 0) {
>> +                ret = fcntl_setfl(rs->lock_fd, rs->open_flags);
>> +                if (ret) {
>> +                    qemu_close(rs->lock_fd);
>> +                    rs->lock_fd = -1;
>> +                }
>> +            }
>> +        }
>>     }
>> 
>>     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
>> @@ -835,6 +848,14 @@ static int raw_reopen_prepare(BDRVReopenState *state,
>>                 error_setg_errno(errp, errno, "Could not reopen file");
>>                 ret = -1;
>>             }
>> +
>> +            if (s->use_lock) {
>> +                rs->lock_fd = qemu_open(normalized_filename, rs->open_flags);
>> +                if (rs->lock_fd == -1) {
>> +                    error_setg_errno(errp, errno, "Could not reopen file for locking");
>> +                    ret = -1;
>> +                }
>> +            }
>>         }
>>     }
>> 
>> @@ -861,7 +882,11 @@ static void raw_reopen_commit(BDRVReopenState *state)
>>     s->open_flags = rs->open_flags;
>> 
>>     qemu_close(s->fd);
>> +    if (s->use_lock) {
>> +        qemu_close(s->lock_fd);
>> +    }
>>     s->fd = rs->fd;
>> +    s->lock_fd = rs->lock_fd;
>> 
>>     g_free(state->opaque);
>>     state->opaque = NULL;
>> -- 
>> 2.14.2
>>
Fam Zheng March 30, 2018, 6:47 a.m. UTC | #2
On Thu, 03/22 18:39, Kevin Wolf wrote:
> [ Cc: qemu-block ]
> 
> Am 22.03.2018 um 18:20 hat Dion Bosschieter geschrieben:
> > In commit 244a5668106297378391b768e7288eb157616f64 another
> > file descriptor to BDRVRawState is added. When we try to issue the
> > reopen command only s->fd is reopened; lock_fd could still hold an old
> > file descriptor "possibly" pointing to another file.
> > 
> > - change raw_reopen_prepare so it checks use_lock from BDRVRawState and
> > tries to reopen lock_fd accordingly
> > - change raw_reopen_commit so it closes the old lock_fd on use_lock
> > 
> > Signed-off-by: Dion Bosschieter <dionbosschieter@gmail.com>
> 
> bdrv_reopen() is not meant for opening a different file, it is meant to
> change the flags and options of the same file. Do you have a use case
> where you would actually need to switch to a different file?
> 
> As far as I know, lock_fd was specifically introduced _because_ it stays
> the same across reopen, so we don't need a racy release/reacquire pair.
> Fam (CCed) should know more.

I think (I remember we have discussed a bit before) techinically we can do
reopen just well without a separate s->lock_fd.  After all all locks we acquire
are shared lock, so it is possible to handle the lock between old fd and the new
one back and forth freely.

Fam
Fam Zheng March 30, 2018, 6:48 a.m. UTC | #3
On Thu, 03/22 19:08, Dion Bosschieter wrote:
> Yeah I have a use case, before a last sync on a storage migration we suspend a
> VM -> send the last diffs -> mount the new storage server and after that we
> change a symlink -> call reopen -> check if all file descriptors are changed
> before resuming the VM.

What is the point of changing the symlink and checking if FDs are changed?

Fam
diff mbox series

Patch

diff --git a/block/file-posix.c b/block/file-posix.c
index d7fb772c14..16d83fc49e 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -167,6 +167,7 @@  typedef struct BDRVRawState {
 
 typedef struct BDRVRawReopenState {
     int fd;
+    int lock_fd;
     int open_flags;
 } BDRVRawReopenState;
 
@@ -795,6 +796,7 @@  static int raw_reopen_prepare(BDRVReopenState *state,
     raw_parse_flags(state->flags, &rs->open_flags);
 
     rs->fd = -1;
+    rs->lock_fd = -1;
 
     int fcntl_flags = O_APPEND | O_NONBLOCK;
 #ifdef O_NOATIME
@@ -820,6 +822,17 @@  static int raw_reopen_prepare(BDRVReopenState *state,
                 rs->fd = -1;
             }
         }
+
+        if (s->use_lock) {
+            rs->lock_fd = qemu_dup(s->lock_fd);
+            if (rs->lock_fd >= 0) {
+                ret = fcntl_setfl(rs->lock_fd, rs->open_flags);
+                if (ret) {
+                    qemu_close(rs->lock_fd);
+                    rs->lock_fd = -1;
+                }
+            }
+        }
     }
 
     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
@@ -835,6 +848,14 @@  static int raw_reopen_prepare(BDRVReopenState *state,
                 error_setg_errno(errp, errno, "Could not reopen file");
                 ret = -1;
             }
+
+            if (s->use_lock) {
+                rs->lock_fd = qemu_open(normalized_filename, rs->open_flags);
+                if (rs->lock_fd == -1) {
+                    error_setg_errno(errp, errno, "Could not reopen file for locking");
+                    ret = -1;
+                }
+            }
         }
     }
 
@@ -861,7 +882,11 @@  static void raw_reopen_commit(BDRVReopenState *state)
     s->open_flags = rs->open_flags;
 
     qemu_close(s->fd);
+    if (s->use_lock) {
+        qemu_close(s->lock_fd);
+    }
     s->fd = rs->fd;
+    s->lock_fd = rs->lock_fd;
 
     g_free(state->opaque);
     state->opaque = NULL;