Re: [Qemu-devel] [PATCH v8 16/16] block/io_uring: adds fd registration

[Stefan Hajnoczi <stefanha@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 4130 bytes --]

On Tue, Jul 30, 2019 at 11:04:41PM +0530, Aarushi Mehta wrote:

I'm concerned about file descriptor leaks.  fd_array[] keeps file
descriptors basically forever, even after the file is no longer in use
by the rest of QEMU.  There needs to be a call to unregister whenever a
file is closed elsewhere in QEMU.  For benchmarking and experimentation
the current code is okay, but for production usage the leak must be
prevented.

> +/**
> + * luring_fd_register:
> + *
> + * Register and unregisters file descriptors, see luring_fd_lookup
> + */
> +static int luring_fd_register(struct io_uring *ring, LuringFd *fd_reg, int fd)
> +{
> +    int ret, nr;
> +    GHashTable *lookup = fd_reg->fd_lookup;
> +    nr = g_hash_table_size(lookup);
> +
> +    /* Unregister */
> +    if (!fd) {
> +        ret = io_uring_unregister_files(ring);
> +        g_hash_table_remove_all(lookup);

Is it correct to clear the hash table be cleared if there was an error?

> +        return ret;
> +    }

Please make unregistering all files a separate function.  It's not
necessary to overload this function since this is a completely separate
operation.

> +
> +    /* If adding new, API requires older registrations to be removed */
> +    if (nr) {
> +        io_uring_unregister_files(ring);
> +    }
> +
> +    fd_reg->fd_array = g_realloc_n(fd_reg->fd_array, nr + 1, sizeof(int));
> +    fd_reg->fd_array[nr] = fd;
> +    fd_reg->fd_index = g_realloc_n(fd_reg->fd_index, nr + 1, sizeof(int));
> +    fd_reg->fd_index[nr] = nr;
> +
> +    g_hash_table_insert(lookup, &fd_reg->fd_array[nr], &fd_reg->fd_index[nr]);

fd_index[] is not necessary, you can cast nr to a gpointer instead to
store the data directly inside GHashTable:

  g_hash_table_insert(lookup, &fd_reg->fd_array[nr],
                      GINT_TO_POINTER(nr));

The hash table accesses can be made slightly more efficient by avoiding
the pointer dereference for keys as well:

  g_hash_table_insert(lookup, GINT_TO_POINTER(fd),
                      GINT_TO_POINTER(nr));

In this case fd_array[] is only used for the io_uring_register_files()
call and nothing else.  Remember to switch to g_direct_equal() and
g_direct_hash() in g_hash_table_new_full() if you make the key a direct
gpointer.

> +    trace_luring_fd_register(fd, nr);
> +    return io_uring_register_files(ring, fd_reg->fd_array, nr + 1);
> +}
> +
> +/**
> + * luring_fd_lookup:
> + *
> + * Used to lookup fd index in registered array at submission time
> + * If the lookup table has not been created or the fd is not in the table,
> + * the fd is registered.
> + *
> + * If registration errors, the hash is cleared and the fd used directly
> + *
> + * Unregistering is done at luring_detach_aio_context
> + */
> +static int luring_fd_lookup(LuringState *s, int fd)
> +{
> +    int *index, ret;
> +    if (!s->fd_reg.fd_lookup) {
> +        s->fd_reg.fd_lookup = g_hash_table_new_full(g_int_hash, g_int_equal,
> +                                                    g_free, g_free);

fd_array[] and fd_index[] are allocated in single allocations for the
entire array, therefore g_free(key) and g_free(value) on individual
elements is undefined behavior and could crash the program.  There
should be no destroy function for them.

Missing g_hash_table_unref() to free fd_lookup.

> +        luring_fd_register(&s->ring, &s->fd_reg, fd);
> +    }
> +    index = g_hash_table_lookup(s->fd_reg.fd_lookup, &fd);
> +
> +    if (!index) {
> +        ret = luring_fd_register(&s->ring, &s->fd_reg, fd);
> +        if (ret < 0) {
> +            g_hash_table_remove_all(s->fd_reg.fd_lookup);

Why is the hash table cleared and why are fd_array[]/fd_index[] left
behind?

> +            return ret;
> +        }
> +        index = g_hash_table_lookup(s->fd_reg.fd_lookup, &fd);
> +    }
> +    return *index;
> +}

What are the concerns about in-flight requests and how are they
addressed?  For example, if a request is in-flight and another request
wants to add a new fd then io_uring_unregister_files() and
io_uring_register_files() are called while a request is still in-flight.
How does the io_uring kernel code handle this?

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]