Re: [dpdk-dev] [PATCH] eal/windows: fix invalid thread handle

[Dmitry Kozlyuk <dmitry.kozliuk@gmail.com>]

On Thu, 21 May 2020 17:11:12 -0700
Tasnim Bashar <tbashar@mellanox.com> wrote:

> Casting thread ID to handle is not accurate way to get thread handle.
> Need to use OpenThread function to get thread handle from thread ID.
> 
> pthread_setaffinity_np and pthread_getaffinity_np functions
> for Windows are affected because of it.

Good catch.

Side note:
The sooner we move to a mature third-party pthread implementation, the better.

> 
> Signed-off-by: Tasnim Bashar <tbashar@mellanox.com>
> ---
>  lib/librte_eal/windows/include/pthread.h | 40 +++++++++++++++++++++---
>  1 file changed, 35 insertions(+), 5 deletions(-)
> 
> diff --git a/lib/librte_eal/windows/include/pthread.h b/lib/librte_eal/windows/include/pthread.h
> index 0bbed5c3b8..d2a986f8fd 100644
> --- a/lib/librte_eal/windows/include/pthread.h
> +++ b/lib/librte_eal/windows/include/pthread.h
> @@ -18,6 +18,7 @@ extern "C" {
>  
>  #include <windows.h>
>  #include <rte_common.h>
> +#include <rte_windows.h>

Build fails, because <rte_windows.h> doesn't include <rte_log.h>, macros from
which it uses. I'd rather include it there, so that <rte_windows.h> could be
used independently, but you can also add an include here.

If you include <rte_windows.h>, you don't need <windows.h> anymore.

>  
>  #define PTHREAD_BARRIER_SERIAL_THREAD TRUE
>  
> @@ -50,7 +51,19 @@ typedef SYNCHRONIZATION_BARRIER pthread_barrier_t;
>  static inline int
>  eal_set_thread_affinity_mask(pthread_t threadid, unsigned long *cpuset)
>  {
> -	SetThreadAffinityMask((HANDLE) threadid, *cpuset);
> +	HANDLE thread_handle = OpenThread(THREAD_ALL_ACCESS, FALSE, threadid);
> +	if (thread_handle == NULL) {
> +		RTE_LOG_WIN32_ERR("OpenThread()");
> +		return -1;
> +	}
> +
> +	DWORD ret = SetThreadAffinityMask(thread_handle, *cpuset);

GetThreadAffinityMask() and SetThreadAffinityMask() operate on DWORD_PTR (8
bytes), not DWORD (4 byte). This applies to all usages of these functions in
the file and also to the type of "cpuset" parameter: it should be either "long
long *" (as rte_cpuset_t is declared) or just "rte_cpuset_t *".

Also, DPDK coding style suggests declaring variables at the start of the block
(i.e. function here and below).

> +	if (ret == 0) {
> +		RTE_LOG_WIN32_ERR("SetThreadAffinityMask()");
> +		CloseHandle(thread_handle);
> +		return -1;
> +	}
> +	CloseHandle(thread_handle);
>  	return 0;
>  }
>  
> @@ -60,12 +73,29 @@ eal_get_thread_affinity_mask(pthread_t threadid, unsigned long *cpuset)
>  	/* Workaround for the lack of a GetThreadAffinityMask()
>  	 *API in Windows
>  	 */
> -		/* obtain previous mask by setting dummy mask */
> -	DWORD dwprevaffinitymask =
> -		SetThreadAffinityMask((HANDLE) threadid, 0x1);

Loss of data for the reason described above: here a 64-bit mask is
stored in a 32-bit variable, so later the wrong value of affinity will be
restored for cores 32 and on.

> +	HANDLE thread_handle = OpenThread(THREAD_ALL_ACCESS, FALSE, threadid);
> +	if (thread_handle == NULL) {
> +		RTE_LOG_WIN32_ERR("OpenThread()");
> +		return -1;
> +	}
> +
> +	/* obtain previous mask by setting dummy mask */
> +	DWORD dwprevaffinitymask = SetThreadAffinityMask(thread_handle, 0x1);
> +	if (dwprevaffinitymask == 0) {
> +		RTE_LOG_WIN32_ERR("SetThreadAffinityMask()");
> +		CloseHandle(thread_handle);
> +		return -1;
> +	}
> +
>  	/* set it back! */
> -	SetThreadAffinityMask((HANDLE) threadid, dwprevaffinitymask);
> +	DWORD ret = SetThreadAffinityMask(thread_handle, dwprevaffinitymask);
> +	if (ret == 0) {
> +		RTE_LOG_WIN32_ERR("SetThreadAffinityMask()");
> +		CloseHandle(thread_handle);
> +		return -1;
> +	}
>  	*cpuset = dwprevaffinitymask;
> +	CloseHandle(thread_handle);
>  	return 0;
>  }
>  

--
Dmitry Kozlyuk