Monitoring an SQLite database with QFileSystemWatcher does not work as expected

[l3u_]

Hi :-)

I'm working on a program using an SQLite database. I'd like to add an error handler for the case that the database is manupulated (renamed, deleted, changed) outside of my program while it's running.

I thought I could simply add a QFileSystemWatcher to monitor the database file and block it's "changed" signal each time my program modifies the database, so that it would only be emitted if the file was changed externally – but this does not seem to work as expected.

Here are some code fragments explaining what's going on:

The QFileSystemWatcher is setup like this:

Database::Database(QObject *parent) : QObject(parent)
{
    ...
    m_dbFileMonitor = new QFileSystemWatcher(this);
    connect(m_dbFileMonitor, &QFileSystemWatcher::fileChanged,
            this, &Database::dbFileChangedExternally);
}

void Database::connectDatabase(const QString &dbFile)
{
    ...
    m_dbFileMonitor->addPath(dbFile);
}

void Database::dbFileChangedExternally(const QString&)
{
    qDebug() << "file changed";
}

Here's some function changing the database:

void Database::registerPlayers(const QString &name)
{
    qDebug() << "blocking signals";
    m_dbFileMonitor->blockSignals(true);
    m_db.transaction();
    QSqlQuery query(m_db);
    query.prepare(QString::fromUtf8("INSERT INTO players(id, name) VALUES(NULL, ?)"));
    query.bindValue(0, name);
    query.exec();
    m_db.commit();
    qDebug() << "unblocking signals";
    m_dbFileMonitor->blockSignals(false);
}

When the function is called, the console output is:

blocking signals
unblocking signals
file changed

So apparently, the QFileSystemWatcher's signal is emitted after QSqlDatabase::commit() has been already called and the signals aren't blocked anymore.

So am I thinking wrong here, or does the actual write operation on the database take place in some asynchronous way? Like I block the file watcher, manipulate the database, unblock it – but the actual writing takes place some short time after, so that it still reports the file to be changed?

How can I solve this? Is there a way to know then the writing will be finished?

Or should I set some bool variable like m_writeInProgress to true when I do something and let the file watcher's changed slot reset it? This way, a change operation would be requested and confirmed by the file watcher – and if something un-requested happens, I know that it wasn't my program?

Or is this simply a complete wrong approach?

Thanks for all help!

[VRonin]

nope, this is just how QFileSystemWatcher works.

QFileSystemWatcher works with polling. It has a timer internally that every second will check the paths you added and store the values. Then, after 1 second, it will check again and if they are different it will trigger the signal.
It's not a chain of signals that gets triggered by m_db.commit(); and so it will not halt Database::registerPlayers. It might take up to 1 second from when control goes back to the event loop to when the signal is triggered.

What you could do is:

• add a int fileChangeCounter=0; private member of Database
• replace connect(m_dbFileMonitor, &QFileSystemWatcher::fileChanged, this, &Database::dbFileChangedExternally); with

connect(m_dbFileMonitor, &QFileSystemWatcher::fileChanged, this, [this]()->void{
if(--fileChangeCounter<0) {
fileChangeCounter=0;
dbFileChangedExternally();
}
}

• delete m_dbFileMonitor->blockSignals(true);
• replace m_dbFileMonitor->blockSignals(false); with ++fileChangeCounter;

But it will still fails if the file will be changed externally in the small span of time between you change the file and the watcher picks it up

P.S.
I'm not 100% sure (I don't know SQLite that well) but I think even a rolled back transaction would change the file

[l3u_]

Thanks for the clarification! So The approach using QFileSystemWatcher is possibly the wrong way at all?!

[JonB]

@VRonin

QFileSystemWatcher works with polling.

I assumed it used the native facilities from the OS....

[Paul Colby]

@JonB said in Monitoring an SQLite database with QFileSystemWatcher does not work as expected:

I assumed it used the native facilities from the OS....

It can, but it depends on the OS. It will, for example, use inotify where possible (and considered reliable), which doesn't require polling, but does poll in lots of other cases.

See QFileSystemWatcherPrivate::createNativeEngine() for some of the engines that might be used.

Cheers.

[mrjj]

Hi
Do you simply what to know if it was changed from outside ?
Other uses cases is to prevent it and using encryption can be one way
https://stackoverflow.com/questions/5669905/sqlite-with-encryption-password-protection/5877130#5877130

[l3u_]

It's not about preventing malicious changes. It's simply about "The user accidentally deletes the database file while the program runs and then it behaves not as expected". I just want to popup a warning "The file has been changed, possibly it's broken now, you better reload it" or "Write access has been removed, can only display from now, not change anymore".

[JonB]

@Paul-Colby
But that doesn't sound like @VRonin 's

QFileSystemWatcher works with polling. It has a timer internally that every second will check the paths you added and store the values. Then, after 1 second, it will check again and if they are different it will trigger the signal.

hence my comment.

[mrjj]

@l3u_
Hi
ok. so its just to know.
What about using Exclusive file locking mode
http://www.sqlite.org/pragma.html#pragma_locking_mode
Your process will then own the file and user cannot delete it.
Or at least it used to work that way. Disclaimer. Not tested/used recently.

[l3u_]

@mrjj Hey, that sounds good :-) Executing the following statements (I suppose there's no QSqlDatabase function for that like QSqlDatabase::transaction()?!)

PRAGMA locking_mode = EXCLUSIVE
BEGIN EXCLUSIVE
COMMIT

actually causes another connection to be read-only, if one wants to change something, one gets an error "database is busy". Nice! This prevents the database to be (accidentally) changed by another process (e. g. two instances of my program opening the same database).

The thing that's left is to know if the user move or deleted the database or changed permissions during runtime …

[mrjj]

@l3u_
Hi so even in EXCLUSIVE mode, the actual file is not locked on OS level?
Like you can rename or move it.
Hmm, i guess i recall incorrectly then.

[l3u_]

@mrjj I can do what I want on filesystem level. It's protected against changes by another SQLite connection though.

[VRonin]

@Paul-Colby said in Monitoring an SQLite database with QFileSystemWatcher does not work as expected:

It can, but it depends on the OS. It will, for example, use inotify where possible

Just to be clearer, even if inotify is used the event will be processed only when control goes back to the event loop so my code above can be considered safer when using inotify as it reduces the delay (from up to 1sec to the time it takes your program to reach the event loop)

[JonB]

@l3u_
I don't really know what you expect to achieve realistically.

It's not clear to me from http://www.sqlite.org/pragma.html#pragma_locking_mode just exactly what they mean by "exclusive". It may only mean that the locks are respected by other SQLite processes, not the OS, and your findings seem to indicate that. Or, it may be that behaviour varies by OS, e.g. locked by OS under Windows but not under Linux.

QFileSystemWatcher is not designed to play along with "I want to know which process made a change, and ignore if it's mine". I don't think you (or it) will have any access to that in information.

The thing that's left is to know if the user move or deleted the database or changed permissions during runtime …

In any case, these may not act as you think they will act, e.g.:

• In Linux certainly (can't remember about Windows), one process can delete a file. If another process has that file open at present, the file is marked for deletion, but not actually deleted till all processes with it open close their connections. (And I don't think your QFileSystemWatcher will notice till it actually gets deleted.)

• When permissions on a file are changed to prevent access, these new permissions have no effect on processes which presently have a handle open on the file, only on processes newly trying to open the file. SQLite presumably holds open handles on the database files all the time, so will carry on as before.

• Even with OS locking of file content, that may well not have any effect on any attempt to delete/rename/change permissions on the file itself.

[l3u_]

@JonB said in Monitoring an SQLite database with QFileSystemWatcher does not work as expected:

@l3u_
I don't really know what you expect to achieve realistically.

The idea is quite simple. The database is not intended to be accessed by multiple processes. So if another process changes something while the program runs, it's possible that it breaks. By setting "locking_mode", I can prevent such changes, as simply the database can't be used by any other connection until it's closed. That's one thing I wanted (and achieved in an elegant way I think, as both my program does handle such a lock now, and e. g. a manual cli connection also refuses to work on the database as long as it's open).

The other thing is to inform the user if e. g. the file has been accidentally deleted while it's open. By now, the program doesn't know it, and on the next change, all data is gone without an error message, which is imo not a desirable behavior. Instead, an error should pop up saying something like "The database file has been deleted, so we reset our program as if we closed it now".

And this is probably possible by monitoring it via a QFileSystemWatcher, isn't it? If I check if it's still there on each change, I know if it was deleted and can handle this.

[VRonin]

Just throwing an idea out there: Qt Creator does this for source files so you should be able to check the source of Qt Creator and find the perfect solution to your problem

[JonB]

@l3u_
As I tried to show in examples, I do not believe you can reliably achieve what you would like to achieve, at least not cross-platform.

[l3u_]

As far as I can see, they also use a QFileSystemWatcher to handle this kind of changes.

[l3u_]

@JonB Apart from the change protection (changes by accident by opening the same database with two program instances), I'm actually only interested in knowing if the file has been deleted during runtime … Should I simply let a QFileSystemWatcher watch it an check it's still there after each change? Or is this the wrong way?

[JonB]

@l3u_
I have said, I'm not convinced it will work in this circumstance, at least under Linux, and I still don't think you've said which OS you're under. But why don't you first test the situation: run up your database Qt app, then go separately try to delete the file and see whether you can? If you can, test to see if QFileSystemWatcher detects it correctly.

I don't think you "let a QFileSystemWatcher watch it an check it's still there after each change". You set up a QFileSystemWatcher, then it will signal you on file delete. Rather than you do any checking at certain periods yourself.