diff --git a/demos/Win32-MinGW/main.c b/demos/Win32-MinGW/main.c
index d2effad35..0830ab149 100644
--- a/demos/Win32-MinGW/main.c
+++ b/demos/Win32-MinGW/main.c
@@ -122,15 +122,16 @@ static const ShellConfig shell_cfg2 = {
 /*
  * Console print server done using synchronous messages. This makes the access
  * to the C printf() thread safe and the print operation atomic among threads.
- * In this example the message is the zero termitated string itself.
+ * In this example the message is the zero terminated string itself.
  */
 static msg_t console_thread(void *arg) {
 
   (void)arg;
   while (!chThdShouldTerminate()) {
-    puts((char *)chMsgWait());
+    Thread *tp = chMsgWait();
+    puts((char *)chMsgGet(tp));
     fflush(stdout);
-    chMsgRelease(RDY_OK);
+    chMsgRelease(tp, RDY_OK);
   }
   return 0;
 }
diff --git a/docs/reports/STM32F103-72-GCC.txt b/docs/reports/STM32F103-72-GCC.txt
index e35cb5270..772e24d45 100644
--- a/docs/reports/STM32F103-72-GCC.txt
+++ b/docs/reports/STM32F103-72-GCC.txt
@@ -98,15 +98,15 @@ Settings: SYSCLK=72, ACR=0x12 (2 wait states)
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 11.1 (Benchmark, messages #1)
---- Score : 248573 msgs/S, 497146 ctxswc/S
+--- Score : 248569 msgs/S, 497138 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 11.2 (Benchmark, messages #2)
---- Score : 201227 msgs/S, 402454 ctxswc/S
+--- Score : 198998 msgs/S, 397996 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 11.3 (Benchmark, messages #3)
---- Score : 201227 msgs/S, 402454 ctxswc/S
+--- Score : 198998 msgs/S, 397996 ctxswc/S
 --- Result: SUCCESS
 ----------------------------------------------------------------------------
 --- Test Case 11.4 (Benchmark, context switch)
diff --git a/docs/reports/kernel.txt b/docs/reports/kernel.txt
index d67788fd3..8f9530559 100644
--- a/docs/reports/kernel.txt
+++ b/docs/reports/kernel.txt
@@ -2,13 +2,13 @@ Platform : PowerPC
 OS Setup : Full kernel
 Compiler : powerpc-eabi-gcc (Sourcery G++ Lite 4.4-79) 4.4.1
 Options  : -O2 -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 11744
+Kernel Size = 11716
 
 Platform : PowerPC
 OS Setup : Full kernel
 Compiler : powerpc-eabi-gcc (Sourcery G++ Lite 4.4-79) 4.4.1
 Options  : -O2 -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 11296
+Kernel Size = 11244
 
 Platform : PowerPC
 OS Setup : Minimal kernel
@@ -20,13 +20,13 @@ Platform : PowerPC
 OS Setup : Full kernel
 Compiler : powerpc-eabi-gcc (Sourcery G++ Lite 4.4-79) 4.4.1
 Options  : -Os -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 10260
+Kernel Size = 10208
 
 Platform : PowerPC
 OS Setup : Full kernel
 Compiler : powerpc-eabi-gcc (Sourcery G++ Lite 4.4-79) 4.4.1
 Options  : -Os -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 9768
+Kernel Size = 9712
 
 Platform : PowerPC
 OS Setup : Minimal kernel
@@ -38,13 +38,13 @@ Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 6172
+Kernel Size = 6124
 
 Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 5672
+Kernel Size = 5624
 
 Platform : ARM Cortex-M3
 OS Setup : Minimal kernel
@@ -56,13 +56,13 @@ Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 5568
+Kernel Size = 5524
 
 Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 5156
+Kernel Size = 5112
 
 Platform : ARM Cortex-M3
 OS Setup : Minimal kernel
@@ -74,13 +74,13 @@ Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 5348
+Kernel Size = 5320
 
 Platform : ARM Cortex-M3
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 4964
+Kernel Size = 4936
 
 Platform : ARM Cortex-M3
 OS Setup : Minimal kernel
@@ -92,13 +92,13 @@ Platform : ARM Cortex-M0
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 5432
+Kernel Size = 5396
 
 Platform : ARM Cortex-M0
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 5236
+Kernel Size = 5196
 
 Platform : ARM Cortex-M0
 OS Setup : Minimal kernel
@@ -110,13 +110,13 @@ Platform : ARM Cortex-M0
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 4992
+Kernel Size = 4956
 
 Platform : ARM Cortex-M0
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 4832
+Kernel Size = 4792
 
 Platform : ARM Cortex-M0
 OS Setup : Minimal kernel
@@ -128,13 +128,13 @@ Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 8988
+Kernel Size = 8940
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 8528
+Kernel Size = 8472
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Minimal kernel
@@ -146,13 +146,13 @@ Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 8344
+Kernel Size = 8288
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 7988
+Kernel Size = 7928
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Minimal kernel
@@ -164,13 +164,13 @@ Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 8704
+Kernel Size = 8676
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 8260
+Kernel Size = 8220
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Minimal kernel
@@ -182,13 +182,13 @@ Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 8052
+Kernel Size = 8016
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 7716
+Kernel Size = 7676
 
 Platform : ARM7TDMI (ARM mode)
 OS Setup : Minimal kernel
@@ -200,13 +200,13 @@ Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=TRUE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 6008
+Kernel Size = 5964
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -DCH_OPTIMIZE_SPEED=FALSE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5800
+Kernel Size = 5752
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Minimal kernel
@@ -218,13 +218,13 @@ Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=TRUE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5580
+Kernel Size = 5536
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -DCH_OPTIMIZE_SPEED=FALSE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5396
+Kernel Size = 5348
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Minimal kernel
@@ -236,13 +236,13 @@ Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=TRUE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5852
+Kernel Size = 5824
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -O2 -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=FALSE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5652
+Kernel Size = 5620
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Minimal kernel
@@ -254,13 +254,13 @@ Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=TRUE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5368
+Kernel Size = 5340
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Full kernel
 Compiler : arm-none-eabi-gcc (GCC) 4.5.2
 Options  : -Os -mthumb -ffixed-r7 -DCH_CURRP_REGISTER_CACHE=\"r7\" -DCH_OPTIMIZE_SPEED=FALSE -DTHUMB -DTHUMB_PRESENT -DTHUMB_NO_INTERWORKING
-Kernel Size = 5212
+Kernel Size = 5180
 
 Platform : ARM7TDMI (THUMB mode)
 OS Setup : Minimal kernel
@@ -272,13 +272,13 @@ Platform : MSP430
 OS Setup : Full kernel
 Compiler : msp430-gcc (GCC) 3.2.3
 Options  : -O2 -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 5884
+Kernel Size = 5840
 
 Platform : MSP430
 OS Setup : Full kernel
 Compiler : msp430-gcc (GCC) 3.2.3
 Options  : -O2 -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 5476
+Kernel Size = 5428
 
 Platform : MSP430
 OS Setup : Minimal kernel
@@ -290,13 +290,13 @@ Platform : MSP430
 OS Setup : Full kernel
 Compiler : msp430-gcc (GCC) 3.2.3
 Options  : -Os -DCH_OPTIMIZE_SPEED=TRUE
-Kernel Size = 5820
+Kernel Size = 5776
 
 Platform : MSP430
 OS Setup : Full kernel
 Compiler : msp430-gcc (GCC) 3.2.3
 Options  : -Os -DCH_OPTIMIZE_SPEED=FALSE
-Kernel Size = 5440
+Kernel Size = 5392
 
 Platform : MSP430
 OS Setup : Minimal kernel
diff --git a/os/kernel/include/chmsg.h b/os/kernel/include/chmsg.h
index 438da021f..b2e8ef51e 100644
--- a/os/kernel/include/chmsg.h
+++ b/os/kernel/include/chmsg.h
@@ -39,20 +39,47 @@
         ((tp)->p_msgqueue.p_next != (Thread *)&(tp)->p_msgqueue)
 
 /**
- * @brief   Returns the first message in the queue.
+ * @brief   Returns the message carried by the specified thread.
+ * @pre     This function must be invoked immediately after exiting a call
+ *          to @p chMsgWait().
  *
- * @iclass
+ * @param[in] tp        pointer to the thread
+ * @return              The message carried by the sender.
+ *
+ * @api
  */
-#define chMsgGetI(tp) \
-        ((tp)->p_msgqueue.p_next->p_msg)
+#define chMsgGet(tp) ((tp)->p_msg)
+
+/**
+ * @brief   Returns the message carried by the specified thread.
+ * @pre     This function must be invoked immediately after exiting a call
+ *          to @p chMsgWait().
+ *
+ * @param[in] tp        pointer to the thread
+ * @return              The message carried by the sender.
+ *
+ * @sclass
+ */
+#define chMsgGetS(tp) ((tp)->p_msg)
+
+/**
+ * @brief   Releases the thread waiting on top of the messages queue.
+ * @pre     Invoke this function only after a message has been received
+ *          using @p chMsgWait().
+ *
+ * @param[in] tp        pointer to the thread
+ * @param[in] msg       message to be returned to the sender
+ *
+ * @sclass
+ */
+#define chMsgReleaseS(tp, msg) chSchWakeupS(tp, msg)
 
 #ifdef __cplusplus
 extern "C" {
 #endif
   msg_t chMsgSend(Thread *tp, msg_t msg);
-  msg_t chMsgWait(void);
-  msg_t chMsgGet(void);
-  void chMsgRelease(msg_t msg);
+  Thread * chMsgWait(void);
+  void chMsgRelease(Thread *tp, msg_t msg);
 #ifdef __cplusplus
 }
 #endif
diff --git a/os/kernel/include/chthreads.h b/os/kernel/include/chthreads.h
index c22255ad0..f6ed23d1d 100644
--- a/os/kernel/include/chthreads.h
+++ b/os/kernel/include/chthreads.h
@@ -176,12 +176,14 @@ struct Thread {
 #define THD_STATE_WTOREVT       8
 /** @brief Thread state: Waiting in @p chEvtWaitAllTimeout().*/
 #define THD_STATE_WTANDEVT      9
-/** @brief Thread state: Waiting in @p chMsgSend().*/
-#define THD_STATE_SNDMSG        10
+/** @brief Thread state: Waiting in @p chMsgSend() (queued).*/
+#define THD_STATE_SNDMSGQ       10
+/** @brief Thread state: Waiting in @p chMsgSend() (not queued).*/
+#define THD_STATE_SNDMSG        11
 /** @brief Thread state: Waiting in @p chMsgWait().*/
-#define THD_STATE_WTMSG         11
+#define THD_STATE_WTMSG         12
 /** @brief Thread state: After termination.*/
-#define THD_STATE_FINAL         12
+#define THD_STATE_FINAL         13
 
 /*
  * Various flags into the thread p_flags field.
@@ -242,7 +244,7 @@ extern "C" {
  * @note    This function is only available when the
  *          @p CH_DBG_THREADS_PROFILING configuration option is enabled.
  *
- * @param[in] tp        the pointer to the thread
+ * @param[in] tp        pointer to the thread
  *
  * @api
  */
@@ -258,7 +260,7 @@ extern "C" {
 /**
  * @brief   Verifies if the specified thread is in the @p THD_STATE_FINAL state.
  *
- * @param[in] tp        the pointer to the thread
+ * @param[in] tp        pointer to the thread
  * @retval TRUE         thread terminated.
  * @retval FALSE        thread not terminated.
  *
@@ -279,7 +281,7 @@ extern "C" {
 /**
  * @brief   Resumes a thread created with @p chThdInit().
  *
- * @param[in] tp        the pointer to the thread
+ * @param[in] tp        pointer to the thread
  *
  * @iclass
  */
@@ -305,7 +307,7 @@ extern "C" {
  *          system clock.
  * @note    The maximum specified value is implementation dependent.
  *
- * @param[in] sec       the time in seconds
+ * @param[in] sec       time in seconds
  *
  * @api
  */
@@ -318,7 +320,7 @@ extern "C" {
  *          system clock.
  * @note    The maximum specified value is implementation dependent.
  *
- * @param[in] msec      the time in milliseconds
+ * @param[in] msec      time in milliseconds
  *
  * @api
  */
@@ -331,7 +333,7 @@ extern "C" {
  *          system clock.
  * @note    The maximum specified value is implementation dependent.
  *
- * @param[in] usec      the time in microseconds
+ * @param[in] usec      time in microseconds
  *
  * @api
  */
diff --git a/os/kernel/src/chmsg.c b/os/kernel/src/chmsg.c
index c3b4848b0..d4199bbb0 100644
--- a/os/kernel/src/chmsg.c
+++ b/os/kernel/src/chmsg.c
@@ -75,7 +75,7 @@ msg_t chMsgSend(Thread *tp, msg_t msg) {
   msg_insert(ctp, &tp->p_msgqueue);
   if (tp->p_state == THD_STATE_WTMSG)
     chSchReadyI(tp);
-  chSchGoSleepS(THD_STATE_SNDMSG);
+  chSchGoSleepS(THD_STATE_SNDMSGQ);
   msg = ctp->p_u.rdymsg;
   chSysUnlock();
   return msg;
@@ -83,69 +83,46 @@ msg_t chMsgSend(Thread *tp, msg_t msg) {
 
 /**
  * @brief   Suspends the thread and waits for an incoming message.
- * @post    After receiving a message the function @p chMsgRelease() must be
- *          invoked in order to acknowledge the reception and send the answer.
+ * @post    After receiving a message the function @p chMsgGet() must be
+ *          called in order to retrieve the message and then @p chMsgRelease()
+ *          must be invoked in order to acknowledge the reception and send
+ *          the answer.
  * @note    If the message is a pointer then you can assume that the data
  *          pointed by the message is stable until you invoke @p chMsgRelease()
  *          because the sending thread is suspended until then.
  *
- * @return              The message.
+ * @return              A reference to the thread carrying the message.
  *
  * @api
  */
-msg_t chMsgWait(void) {
-  msg_t msg;
+Thread *chMsgWait(void) {
+  Thread *tp;
 
   chSysLock();
   if (!chMsgIsPendingI(currp))
     chSchGoSleepS(THD_STATE_WTMSG);
-#if defined(CH_ARCHITECTURE_STM8)
-  msg = chMsgGetI((volatile Thread *)currp); /* Temporary hack.*/
-#else
-  msg = chMsgGetI(currp);
-#endif
+  tp = fifo_remove(&currp->p_msgqueue);
+  tp->p_state = THD_STATE_SNDMSG;
   chSysUnlock();
-  return msg;
-}
-
-/**
- * @brief   Returns the next message in the queue.
- * @post    After receiving a message the function @p chMsgRelease() must be
- *          invoked in order to acknowledge the reception and send the answer.
- * @note    If the message is a pointer then you can assume that the data
- *          pointed by the message is stable until you invoke @p chMsgRelease()
- *          because the sending thread is suspended until then.
- *
- * @return              The message.
- * @retval 0            if the queue is empty.
- *
- * @api
- */
-msg_t chMsgGet(void) {
-  msg_t msg;
-
-  chSysLock();
-  msg = chMsgIsPendingI(currp) ? chMsgGetI(currp) : (msg_t)NULL;
-  chSysUnlock();
-  return msg;
+  return tp;
 }
 
 /**
  * @brief   Releases the thread waiting on top of the messages queue.
  * @pre     Invoke this function only after a message has been received
- *          using @p chMsgWait() or @p chMsgGet().
+ *          using @p chMsgWait().
  *
- * @param[in] msg       the message returned to the message sender
+ * @param[in] tp        pointer to the thread
+ * @param[in] msg       message to be returned to the sender
  *
  * @api
  */
-void chMsgRelease(msg_t msg) {
+void chMsgRelease(Thread *tp, msg_t msg) {
 
   chSysLock();
-  chDbgAssert(chMsgIsPendingI(currp),
-              "chMsgRelease(), #1",
-              "no message pending");
-  chSchWakeupS(fifo_remove(&currp->p_msgqueue), msg);
+  chDbgAssert(tp->p_state == THD_STATE_SNDMSG,
+              "chMsgRelease(), #1", "invalid state");
+  chMsgReleaseS(tp, msg);
   chSysUnlock();
 }
 
diff --git a/os/kernel/src/chmtx.c b/os/kernel/src/chmtx.c
index b84c4d57e..826ef27fe 100644
--- a/os/kernel/src/chmtx.c
+++ b/os/kernel/src/chmtx.c
@@ -141,7 +141,7 @@ void chMtxLockS(Mutex *mp) {
       case THD_STATE_WTSEM:
 #endif
 #if CH_USE_MESSAGES_PRIORITY
-      case THD_STATE_SNDMSG:
+      case THD_STATE_SNDMSGQ:
 #endif
         /* Re-enqueues tp with its new priority on the queue.*/
         prio_insert(dequeue(tp), (ThreadsQueue *)tp->p_u.wtobjp);
diff --git a/readme.txt b/readme.txt
index 4ca9da65d..9f18bdb52 100644
--- a/readme.txt
+++ b/readme.txt
@@ -96,6 +96,8 @@
   driver.
 - CHANGE: Swapped the numeric values of the TIME_IMMEDIATE and TIME_INFINITE
   constants. Fixed the relative documentation in various places.
+- CHANGE: Slightly modified the messages API in order to allow the
+  processing of multiple messages at the same time.
 
 *** 2.1.8 ***
 - FIX: Fixed error in STM32 ADC driver macro names (bug 3160306)(backported
diff --git a/test/testbmk.c b/test/testbmk.c
index 0f38eb78b..7e7a714df 100644
--- a/test/testbmk.c
+++ b/test/testbmk.c
@@ -65,11 +65,14 @@ static Mutex mtx1;
 #endif
 
 static msg_t thread1(void *p) {
+  Thread *tp;
   msg_t msg;
 
   (void)p;
   do {
-    chMsgRelease(msg = chMsgWait());
+    tp = chMsgWait();
+    msg = chMsgGet(tp);
+    chMsgRelease(tp, msg);
   } while (msg);
   return 0;
 }
diff --git a/test/testmsg.c b/test/testmsg.c
index 881db167f..14b6d8186 100644
--- a/test/testmsg.c
+++ b/test/testmsg.c
@@ -64,11 +64,11 @@ static msg_t thread(void *p) {
   chMsgSend(p, 'A');
   chMsgSend(p, 'B');
   chMsgSend(p, 'C');
-  chMsgSend(p, 'D');
   return 0;
 }
 
 static void msg1_execute(void) {
+  Thread *tp;
   msg_t msg;
 
   /*
@@ -76,29 +76,19 @@ static void msg1_execute(void) {
    */
   threads[0] = chThdCreateStatic(wa[0], WA_SIZE, chThdGetPriority() + 1,
                                  thread, chThdSelf());
-  chMsgRelease(msg = chMsgWait());
+  tp = chMsgWait();
+  msg = chMsgGet(tp);
+  chMsgRelease(tp, msg);
   test_emit_token(msg);
-  chMsgRelease(msg = chMsgWait());
+  tp = chMsgWait();
+  msg = chMsgGet(tp);
+  chMsgRelease(tp, msg);
   test_emit_token(msg);
-  chMsgRelease(msg = chMsgWait());
+  tp = chMsgWait();
+  msg = chMsgGet(tp);
+  chMsgRelease(tp, msg);
   test_emit_token(msg);
   test_assert_sequence(1, "ABC");
-
-  /*
-   * Testing message fetch using chMsgGet().
-   * Note, the following is valid because the sender has higher priority than
-   * the receiver.
-   */
-  msg = chMsgGet();
-  test_assert(1, msg != 0, "no message");
-  chMsgRelease(0);
-  test_assert(2, msg == 'D', "wrong message");
-
-  /*
-   * Must not have pending messages.
-   */
-  msg = chMsgGet();
-  test_assert(3, msg == 0, "unknown message");
 }
 
 ROMCONST struct testcase testmsg1 = {