Re: ntdll: Support receiving server reply data to memory with write watch.

[Alexandre Julliard]

Jacek Caban  writes:
> @@ -256,6 +256,7 @@ static void read_reply_data( void *buffer, size_t size )
>          if (!ret) break;
>          if (errno == EINTR) continue;
>          if (errno == EPIPE) break;
> -
>    if (errno == EFAULT && virtual_uninterrupted_write_memory( buffer, NULL, size ) >= size) continue;
>    server_protocol_perror("read");

You can't enter a critical section in the middle of a server call.

ntdll: Support receiving server reply data to memory with write watch.

[Jacek Caban]

On 03.08.2017 09:48, Alexandre Julliard wrote:
> Jacek Caban  writes:
> > @@ -256,6 +256,7 @@ static void read_reply_data( void *buffer, size_t size )
> >          if (!ret) break;
> >          if (errno == EINTR) continue;
> >          if (errno == EPIPE) break;
> > -
> >    if (errno == EFAULT && virtual_uninterrupted_write_memory( buffer, NULL, size ) >= size) continue;
> >    server_protocol_perror("read");
>
> You can't enter a critical section in the middle of a server call.

Oh, right, we could need to do another server call there. Unless I'm
missing something, it should work fine for platforms that have optimized
CS implementation (Linux and Mac) as they don't do server calls. We
could limit effect of the patch to those platform, but that would be ugly.

Could we just use pthread mutexes for
server_enter_uninterrupted_section/server_leave_uninterrupted_section
instead of critical sections?

Thanks,
Jacek

ntdll: Support receiving server reply data to memory with write watch.

[Alexandre Julliard]

Jacek Caban  writes:
> On 03.08.2017 09:48, Alexandre Julliard wrote:
> > Jacek Caban  writes:
> > > @@ -256,6 +256,7 @@ static void read_reply_data( void *buffer, size_t size )
> > >          if (!ret) break;
> > >          if (errno == EINTR) continue;
> > >          if (errno == EPIPE) break;
> > > -
> > >    if (errno == EFAULT && virtual_uninterrupted_write_memory( buffer, NULL, size ) >= size) continue;
> > >    server_protocol_perror("read");
> >
> > You can't enter a critical section in the middle of a server call.
>
> Oh, right, we could need to do another server call there. Unless I'm
> missing something, it should work fine for platforms that have optimized
> CS implementation (Linux and Mac) as they don't do server calls. We
> could limit effect of the patch to those platform, but that would be ugly.
>
> Could we just use pthread mutexes for
> server_enter_uninterrupted_section/server_leave_uninterrupted_section
> instead of critical sections?

I don't think you should be doing that sort of thing inside server calls
at all. Some sort of lock/unlock_buffer methods that could be wrapped
around server calls where necessary would probably work better.