[Spice-devel] [PATCH spice-server 03/13] char-device: Allocate all write buffer in a single block

Frediano Ziglio fziglio at redhat.com
Fri Aug 2 11:07:07 UTC 2019


> 
> Hi,
> 
> On Thu, May 30, 2019 at 03:22:44PM +0100, Frediano Ziglio wrote:
> > There are no much data other than the buffer, reduce the
> > allocations.
> > 
> > Signed-off-by: Frediano Ziglio <fziglio at redhat.com>
> > ---
> >  server/char-device.c | 27 +++++++++++++++------------
> >  server/char-device.h |  2 +-
> >  2 files changed, 16 insertions(+), 13 deletions(-)
> > 
> > diff --git a/server/char-device.c b/server/char-device.c
> > index 89581ea42..059e8e5f6 100644
> > --- a/server/char-device.c
> > +++ b/server/char-device.c
> > @@ -149,11 +149,9 @@ red_char_device_remove_client(RedCharDevice *dev,
> > RedClient *client)
> >  
> >  static void red_char_device_write_buffer_free(RedCharDeviceWriteBuffer
> >  *buf)
> >  {
> > -    if (buf == NULL)
> > -        return;
> > -
> > -    g_free(buf->buf);
> > -    g_free(buf);
> > +    if (buf) {
> > +        g_free(buf->priv);
> > +    }
> >  }
> >  
> >  static void write_buffers_queue_free(GQueue *write_queue)
> > @@ -542,22 +540,27 @@ red_char_device_write_buffer_get(RedCharDevice *dev,
> > RedClient *client, int size
> >      ret = g_queue_pop_tail(&dev->priv->write_bufs_pool);
> >      if (ret) {
> >          dev->priv->cur_pool_size -= ret->buf_size;
> > -    } else {
> > +        if (ret->buf_size < size) {
> > +            spice_assert(!spice_extra_checks || ret->priv->refs == 1);
> 
> Weird we didn't check this before...
> 

Does not hurt, disabled by default

> > +            red_char_device_write_buffer_free(ret);
> > +            ret = NULL;
> 
> I think that in the kernel they do this kind of _free() function
> to return NULL just to make this kind of function call nicer,
> that is,
>     
>     ret = red_char_device_write_buffer_free(ret);
> 

That's weird

> while we often use glib's
> 
>     g_clear_pointer(&ret, red_char_device_write_buffer_free);
> 

Honestly g_clear_pointer implementation is quite bad, really
type unsafe. You probably can pass a string as second argument
and compiler won't complain

> ... just a silly comment :)
> 
> > +        }
> > +    }
> > +    if (ret == NULL) {
> >          struct RedCharDeviceWriteBufferFull {
> > -            RedCharDeviceWriteBuffer buffer;
> >              RedCharDeviceWriteBufferPrivate priv;
> > +            RedCharDeviceWriteBuffer buffer;
> >          } *write_buf;
> > -        write_buf = g_new0(struct RedCharDeviceWriteBufferFull, 1);
> > +        write_buf = g_malloc(sizeof(struct RedCharDeviceWriteBufferFull) +
> > size);
> > +        memset(write_buf, 0, sizeof(*write_buf));
> 
> g_malloc0?
> 

It would clear the entire buffer, this way the data are not cleared.
(Don't underestimate memory copies and reset!)

> > +        write_buf->priv.refs = 1;
> 
> Weird that this was not being set before...
> 

It was set in a different path, but was set (maybe putting in cache or
extracting).

> >          ret = &write_buf->buffer;
> > +        ret->buf_size = size;
> >          ret->priv = &write_buf->priv;
> >      }
> >  
> >      spice_assert(!ret->buf_used);
> >  
> > -    if (ret->buf_size < size) {
> > -        ret->buf = g_realloc(ret->buf, size);
> > -        ret->buf_size = size;
> > -    }
> 
> Looks fine,
> Acked-by: Victor Toso <victortoso at redhat.com>
> 
> >      ret->priv->origin = origin;
> >  
> >      if (origin == WRITE_BUFFER_ORIGIN_CLIENT) {
> > diff --git a/server/char-device.h b/server/char-device.h
> > index 7d3ad8b3a..5327c25e5 100644
> > --- a/server/char-device.h
> > +++ b/server/char-device.h
> > @@ -151,11 +151,11 @@ GType red_char_device_get_type(void) G_GNUC_CONST;
> >  /* buffer that is used for writing to the device */
> >  typedef struct RedCharDeviceWriteBufferPrivate
> >  RedCharDeviceWriteBufferPrivate;
> >  typedef struct RedCharDeviceWriteBuffer {
> > -    uint8_t *buf;
> >      uint32_t buf_size;
> >      uint32_t buf_used;
> >  
> >      RedCharDeviceWriteBufferPrivate *priv;
> > +    uint8_t buf[0];
> >  } RedCharDeviceWriteBuffer;
> >  
> >  void red_char_device_reset_dev_instance(RedCharDevice *dev,

Frediano


More information about the Spice-devel mailing list