Re: [v3 PATCH 1/2] dwrite: Protect cached fontface list when accessed from multiple threads

[Sebastian Lackner]

On 10.08.2017 08:45, Nikolay Sivov wrote:
> Based on patch by Anton Romanov
> Signed-off-by: Nikolay Sivov
>
> dlls/dwrite/dwrite_private.h |  6 ++++--
>  dlls/dwrite/font.c           | 24 ++++++++++++---------
>  dlls/dwrite/main.c           | 50 +++++++++++++++++++++++++++++++++-----------
>  3 files changed, 56 insertions(+), 24 deletions(-)

Isn't it also necessary to protect factory_cache_fontface()?

By the way: An alternative method to implement proper synchronization (which does
not require locking critical sections before InterlockedDecrement) would be to do
something like:

In the destructor:

--- snip ---
    if (!ref)
    {
        EnterCriticalSection(...);
        list_remove(...);
        LeaveCriticalSection(...);

...
    }
--- snip ---

And when querying cached font faces:

--- snip ---
    EnterCriticalSection(...);
    LIST_FOR_EACH_ENTRY(...)
    {
        ...

/* the following code could also be hidden in QueryInterface */
        if (InterlockedIncrement( &obj->refcount ) == 1)
        {
            InterlockedDecrement( &obj->refcount );
            continue;
        }

/* return it */
    }
    LeaveCriticalSection(...);
--- snip ---

This is also what is done in Wines threadpool implementation [1], and has the
advantage that the CS is not needed for every Release, only when the object is
really destroyed. But in the end it is a matter of taste, for rpcrt4 Jacek
also picked the "hold CS during InterlockedDecrement" method.

[1] https://source.winehq.org/git/wine.git/blob/HEAD:/dlls/ntdll/threadpool.c#l2...

Best regards,
Sebastian

[v3 PATCH 1/2] dwrite: Protect cached fontface list when accessed from multiple threads

[Nikolay Sivov]

On 10.08.2017 11:14, Sebastian Lackner wrote:
> On 10.08.2017 08:45, Nikolay Sivov wrote:
> > Based on patch by Anton Romanov
> > Signed-off-by: Nikolay Sivov
> >
> > dlls/dwrite/dwrite_private.h |  6 ++++--
> >  dlls/dwrite/font.c           | 24 ++++++++++++---------
> >  dlls/dwrite/main.c           | 50 +++++++++++++++++++++++++++++++++-----------
> >  3 files changed, 56 insertions(+), 24 deletions(-)
>
> Isn't it also necessary to protect factory_cache_fontface()?

It is.
> By the way: An alternative method to implement proper synchronization (which does
> not require locking critical sections before InterlockedDecrement) would be to do
> something like:
>
> In the destructor:
>
> --- snip ---
>     if (!ref)
>     {
>         EnterCriticalSection(...);
>         list_remove(...);
>         LeaveCriticalSection(...);
>     ...
> }
>
> --- snip ---

Yes, that's better I guess, thanks.
> And when querying cached font faces:
>
> --- snip ---
>     EnterCriticalSection(...);
>     LIST_FOR_EACH_ENTRY(...)
>     {
>         ...
>     /* the following code could also be hidden in QueryInterface */
>     if (InterlockedIncrement( &obj->refcount ) == 1)
>     {
>         InterlockedDecrement( &obj->refcount );
>         continue;
>     }
>
>     /* return it */
> }
> LeaveCriticalSection(...);
>
> --- snip ---

In my case, to avoid exposing structure, it's enough to AddRef() == 1,
and release otherwise I think. It's not important to decrement when
we're in a process of destroying the object.
> This is also what is done in Wines threadpool implementation [1], and has the
> advantage that the CS is not needed for every Release, only when the object is
> really destroyed. But in the end it is a matter of taste, for rpcrt4 Jacek
> also picked the "hold CS during InterlockedDecrement" method.
>
> [1] https://source.winehq.org/git/wine.git/blob/HEAD:/dlls/ntdll/threadpool.c#l2...
>
> Best regards,
> Sebastian

[v3 PATCH 1/2] dwrite: Protect cached fontface list when accessed from multiple threads

[Anton Romanov]

> --- snip ---
>     if (!ref)
>     {
>         EnterCriticalSection(...);
>         list_remove(...);
>         LeaveCriticalSection(...);
>     ...
> }
>
> --- snip ---
>
> And when querying cached font faces:
>
> --- snip ---
>     EnterCriticalSection(...);
>     LIST_FOR_EACH_ENTRY(...)
>     {
>         ...
>     /* the following code could also be hidden in QueryInterface */
>     if (InterlockedIncrement( &obj->refcount ) == 1)
>     {
>         InterlockedDecrement( &obj->refcount );
>         continue;
>     }
>
>     /* return it */
> }
> LeaveCriticalSection(...);
>
> --- snip ---

Doesn't the object just leak in this scenario though ? (assuming
current refcount is 1) :
 ============= ===========
   Thread A     Thread B
 ============= ===========
                enter_cs
                Increment
  Decrement
  ref != Zero
                Decrement
                leave_cs
 ============= ===========